Compare commits

...
6 Commits
Author SHA1 Message Date
VallierDevandClaude Opus 5.5 9906dd4767 Release v0.2.22
Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-10-22 15:14:55 +08:00
Jeff Zhang 9e2198b03d docs: AMD VAAPI verified on the in-process engine; CPU cost figures (#81)
README/README.en/AGENTS: AMD iGPUs (HawkPoint, Cezanne) are accepted on the inproc-av VAAPI engine with the official
Mac viewer; the FFmpeg process pipeline on AMD is not separately verified. inproc-av now covers VAAPI in process;
build headers listed. Thanks @ferstar.
2026-10-21 22:24:42 -07:00
Jeff Zhang 21c01d2e18 inproc-av: native VAAPI H.264/HEVC through DRI3 DMA-BUF, VPP and VCN (#80)
With --features inproc-av on X11 and UU_VIDEO_ENCODER=vaapi, VAAPI runs in process: a DRI3 DMA-BUF of the root
(XShm upload without DRI3) goes through VPP to NV12 and the hardware encoder. SPS/PPS(/VPS) and slice headers are
packed by the host because radeonsi (Mesa 26) does not write them. libva/libva-drm are dlopen'ed; release builds do
not enable inproc-av. The NVENC path is unchanged.

Accepted by the contributor on AMD HawkPoint1 (Mesa 26.2.3), 1920x1200@60, with the official Mac viewer
(HEVC, presets, fps, scaling). Intel iHD and 4K are not verified. Thanks @ferstar.
2026-10-21 15:25:13 -07:00
059a3b1bb8 inproc-av: 4:4:4 on the native NVENC engine (#76, #77)
H.264 High 4:4:4 / HEVC FRExt with chromaFormatIDC 3 and AQ off when 真彩 is requested; GPUs without YUV444 encode fail the capability check and fall back. full_chroma_available() no longer excludes the in-process path. UU_PROBE_CHROMA=444 for the probe.

Co-authored-by: ferstar <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-10-21 08:24:32 -07:00
76941459cc inproc-av: native in-process NVENC reading XShm directly (#74, #75)
Replaces the FFmpeg-based in-process encoder with a native NVENC engine (vendored nv-codec-headers n12.1.14.0, libcuda/libnvidia-encode dlopen'ed at runtime, primary context reuse). Fallback: native NVENC -> FFmpeg NVENC process -> libx264. inproc-av no longer links FFmpeg; VAAPI/CPU always use the process pipeline. Default builds are unchanged.

Co-authored-by: ferstar <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-10-21 07:26:45 -07:00
1fa346fd1f device_capability: list every supported format with its size, not 0x0 (#79)
The official viewer (libstreamer, Mac 4.43) treats a video_codec_capability entry
with width or height 0 as unsupported. The 4.37 Windows host we copied in #49 ran on
a GTX 660 without HEVC/4:4:4, so its 0x0 entries meant exactly that. Evidence by
@ferstar in #77.

Co-authored-by: VallierDev <[email protected]>
Co-authored-by: ferstar <[email protected]>
Co-authored-by: Claude Opus 5.5 <[email protected]>
2026-10-21 21:29:23 +08:00
24 changed files with 7727 additions and 463 deletions

No files matched your search

+2 -2
View File
@@ -94,12 +94,12 @@ jobs:
runs-on: ubuntu-24.04
steps:
- uses: actions/checkout@v4
- name: Install build dependencies (X11 + FFmpeg 6 for the inproc-av feature)
- name: Install build dependencies (X11; inproc-av vendors the NVENC headers; the VAAPI/DRI3 engines need libva/gbm/xcb-dri3 headers, libva itself is dlopen'ed at runtime)
run: |
sudo apt-get update
sudo apt-get install -y build-essential pkg-config \
libx11-dev libxext-dev libxtst-dev libxfixes-dev libgtk-3-dev \
libavcodec-dev libavformat-dev libavdevice-dev libswscale-dev libclang-dev
libva-dev libgbm-dev libx11-xcb-dev libxcb-dri3-dev libxcb-damage0-dev
- uses: dtolnay/rust-toolchain@stable
- uses: Swatinem/rust-cache@v2
- run: cargo test --locked --features inproc-av
+21 -10
View File
@@ -38,9 +38,13 @@ Dependencies (Ubuntu/Debian):
```sh
sudo apt install build-essential pkg-config \
libx11-dev libxext-dev libxtst-dev libxfixes-dev libavcodec-dev libavformat-dev libavdevice-dev libswscale-dev libclang-dev libgtk-3-dev ffmpeg
# --features inproc-av: libavcodec-dev libavformat-dev libavdevice-dev libswscale-dev libclang-dev
# (FFmpeg 6.x only; in-process capture/encoding, Ubuntu 24.04 ships it. FFmpeg >= 7
# fails in ffmpeg-sys-next with "avfft.h file not found", before our build.rs runs)
# --features inproc-av: X11 in-process capture + native NVENC/VAAPI. NVENC headers are vendored and
# libcuda/libnvidia-encode dlopen'ed at runtime (NVENC API 12.1, driver >= 530).
# The VAAPI/DRI3 engines compile against
# libva-dev libgbm-dev libx11-xcb-dev libxcb-dri3-dev libxcb-damage-dev;
# libva/libva-drm themselves are dlopen'ed at runtime (no hard libva dependency)
# --features viewer-ffmpeg: libavcodec-dev libavformat-dev libavdevice-dev libswscale-dev libclang-dev
# (FFmpeg 6.x only. FFmpeg >= 7 fails in ffmpeg-sys-next with "avfft.h file not found")
# --features wayland: libwayland-dev libxkbcommon-dev wl-clipboard
# --features wayland,portal: + libglib2.0-dev libpipewire-0.3-dev libei-dev (libei >= 1.1)
```
@@ -160,7 +164,7 @@ every 60 s.
Diagnostic subcommands: `doctor` (paste-safe report: session, selected backend, engine plugins, uinput, encoders,
default codec, sign-in/service state; booleans and names only), `login [--no-gui]`, `unlock-check` (password on stdin, prints only accepted/rejected),
`desktop-probe [new.png]`, `desktop-video-probe` (`UU_PROBE_FPS`, `UU_PROBE_FRAMES`, `UU_PROBE_QUALITY` 1-5 = viewer bitrate preset, 5 auto; `UU_PROBE_INPUT` fixed video/image sequence, `UU_PROBE_LOOP=1` repeats), `wayland-input-probe`, `clipboard-acceptance`,
`desktop-probe [new.png]`, `desktop-video-probe` (`UU_PROBE_FPS`, `UU_PROBE_FRAMES`, `UU_PROBE_QUALITY` 1-5 = viewer bitrate preset, 5 auto; `UU_PROBE_CHROMA=444` 真彩; `UU_PROBE_INPUT` fixed video/image sequence, `UU_PROBE_LOOP=1` repeats), `wayland-input-probe`, `clipboard-acceptance`,
`portal-acceptance`, `help`. Unknown arguments exit with code 2. Builds up to v0.1.0 did **not** reject them and
started a second full host server instead, so check that a subcommand exists in the binary you run.
@@ -183,7 +187,7 @@ started a second full host server instead, so check that a subcommand exists in
5 Auto, 6 Custom. Resolution: 16/17 pixel size or 5/6, 10 resolution_type (3 = follow remote).
Field 7 `chorma_format` (ChromaFormat 1 = 4:2:0, 3 = 4:4:4) is sent when 真彩 is toggled; ff_capture_setting versions
are cumulative (3 CHROMA_SAMPLING, 4 CUSTOM_BITRATE, 5 DPI_SCALE, 6 HDR). 4:4:4 = NVENC `high444p` / x264 `high444`
with `yuv444p`; VAAPI and inproc stay 4:2:0.
with `yuv444p`; the in-process NVENC engine uses High 4:4:4 / HEVC FRExt (AQ off); VAAPI stays 4:2:0.
Reply `RpcResponse{2 CaptureSettingResponse{}}` (no errors = applied) and send Message 25 CaptureSettingSync
(field 3 = the chroma actually encoded).
- ConnectOptions field 4 `decoder_cap_list` = DecoderCap {1 fps, 2 codec_type (1 H264, 2 H265), 3 width, 4 height,
@@ -275,12 +279,19 @@ started a second full host server instead, so check that a subcommand exists in
`UU_VIDEO_ENCODER=vaapi` selects Intel/AMD H.264 VAAPI; `UU_VAAPI_DEVICE` defaults to
`/dev/dri/renderD128` (render-node permissions required). No automatic VAAPI selection.
Both FFmpeg process and `--features inproc-av` upload NV12 frames to a VAAPI frame pool.
The FFmpeg process pipeline uploads NV12 frames to a VAAPI frame pool. With `--features inproc-av` on X11, VAAPI
H.264/HEVC Main runs in process: a DRI3 DMA-BUF of the root (XShm upload without DRI3) goes through VPP to NV12
and VCN, with SPS/PPS(/VPS) and slice headers packed by the host as radeonsi (Mesa 26) needs; parameters follow
FFmpeg's hevc_vaapi/h264_vaapi on that driver. Accepted on AMD HawkPoint1 (Mesa 26.2.3) and Cezanne
(Mesa 26.2.5; XShm-upload and DRI3 zero-copy probes plus a full HEVC Mac-viewer session) with the official
Mac viewer (presets, fps, scaling); Intel iHD is not verified. Measured on Cezanne + RTX 3050 Ti
(1920x1200, desktop-video-probe, 300 frames): in-process VAAPI ~1.5 ms CPU/frame vs in-process NVENC
~7.1 ms vs libx264 process pipeline ~19 ms.
VBR uses the viewer bitrate preset and CQ-derived minimum QP, no B-frames, async depth 1,
AUD and the existing GOP/VFR timing. Unsupported device/driver/initialization falls back to libx264
with a warning. Accepted on Intel UHD 770 (iHD) with the default FFmpeg process pipeline and the official Mac
viewer (presets, fps, scaling, reconnect; no fallback, ~40% less CPU than libx264). AMD and the `inproc-av`
VAAPI path are not hardware-verified.
viewer (presets, fps, scaling, reconnect; no fallback, ~40% less CPU than libx264). The process pipeline on
AMD is not separately verified (AMD acceptance is the in-process engine above).
### Virtual audio / 虚拟声卡
@@ -321,8 +332,8 @@ This permission can inject into the whole seat, including lock/greeter screens.
Optional VAAPI: `intel-media-driver` (modern Intel) or `libva-mesa-driver` (AMD), `libva-utils`
for `vainfo`. Optional KMS: `libdrm`. Virtual Xfce: `xorg-server-xvfb xorg-xauth xorg-xdpyinfo
xfce4-session xfce4-panel xfdesktop xfwm4 gnome-terminal`.
Rolling Arch FFmpeg may be newer than supported `inproc-av` (FFmpeg 6.x); use the default process
pipeline or build against a separate FFmpeg 6 prefix. Do not downgrade the system FFmpeg.
Rolling Arch FFmpeg may be newer than supported `viewer-ffmpeg` (FFmpeg 6.x); build it against a separate
FFmpeg 6 prefix. Do not downgrade the system FFmpeg. The host's `inproc-av` does not link FFmpeg.
User-bus headless setup and the official Mac checklist: [virtual desktop acceptance](contrib/virtual-desktop/ACCEPTANCE.md).
### Experimental KMS / 实验后端
Generated
+1 -1
View File
@@ -2741,7 +2741,7 @@ checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"
[[package]]
name = "uu-remote-linux"
version = "0.2.21"
version = "0.2.22"
dependencies = [
"async-trait",
"base64 0.22.1",
+9 -4
View File
@@ -1,6 +1,6 @@
[package]
name = "uu-remote-linux"
version = "0.2.21"
version = "0.2.22"
edition = "2021"
default-run = "uuremote-server"
authors = ["uu-remote-linux contributors"]
@@ -30,9 +30,9 @@ kms = []
portal = []
# Build the wayland/portal engines as runtime-loaded libuu-wayland.so / libuu-portal.so (release packages).
engine-plugins = []
# In-process capture and encoding through FFmpeg 6 libraries (libavcodec/libavdevice dev packages).
# Off by default so release binaries do not link the build machine's libav.
inproc-av = ["dep:ffmpeg-next", "dep:x11rb"]
# In-process X11 capture and NVENC encoding: vendored NVENC headers, NVIDIA driver libraries loaded at runtime.
# Other encoders keep the FFmpeg process pipeline. Off by default.
inproc-av = ["dep:x11rb"]
[dependencies]
tokio = { version = "1", features = ["full"] }
@@ -67,3 +67,8 @@ cc = "1.0"
webrtc-dtls = { path = "vendor/webrtc-dtls" }
# Larger data-channel receive buffer (see vendor/webrtc/src/data_channel/mod.rs).
webrtc = { path = "vendor/webrtc" }
[[bench]]
name = "nvenc_pipeline_bench"
harness = false
required-features = ["inproc-av"]
+18 -10
View File
@@ -117,7 +117,7 @@ Switching reuses one video track and pauses input until the new source sends a f
| Feature | Status |
|---|---|
| Switching between multiple monitors | Experimental on native Hyprland Wayland; see above |
| True color (4:4:4) | Supported (v0.1.15): the default FFmpeg pipeline with NVENC / libx264 encodes H.264 High 4:4:4; VAAPI and `inproc-av` stay 4:2:0 and report that to the client |
| True color (4:4:4) | Supported (v0.1.15): the default FFmpeg pipeline with NVENC / libx264 encodes H.264 High 4:4:4; `inproc-av` native NVENC encodes H.264 High 4:4:4 / HEVC FRExt; VAAPI stays 4:2:0 and reports that to the client |
| HDR | Not implemented |
| Super screen / virtual display | Not implemented |
| Privacy screen on the host | Not implemented |
@@ -172,6 +172,8 @@ sudo apt install build-essential pkg-config \
sudo apt install libwayland-dev libxkbcommon-dev wl-clipboard
# optional: GNOME/KDE portal backend
sudo apt install libglib2.0-dev libpipewire-0.3-dev libei-dev
# optional: headers for the inproc-av VAAPI/DRI3 engines (libva itself is loaded at runtime)
sudo apt install libva-dev libgbm-dev libx11-xcb-dev libxcb-dri3-dev libxcb-damage0-dev
```
- A Rust toolchain is required (<https://rustup.rs>).
@@ -187,10 +189,8 @@ cargo build --release --features wayland # + Hyprland
cargo build --release --features wayland,portal # + GNOME/KDE on Wayland
```
The optional `--features inproc-av` (in-process encoding) and the viewer's `viewer-ffmpeg` support **FFmpeg 6.x only**
(`pkg-config --modversion libavcodec` = 60.x). FFmpeg 7+ fails inside the `ffmpeg-sys-next` dependency with
`avfft.h file not found`; its build script runs before this project's, so no clearer check is possible (#63). The default
build and the release packages need neither feature.
The viewer's `viewer-ffmpeg` feature supports **FFmpeg 6.x only** (`pkg-config --modversion libavcodec` = 60.x); FFmpeg 7+ fails inside `ffmpeg-sys-next` with `avfft.h file not found` (#63).
The host's optional `--features inproc-av` encodes with hardware inside the process on X11. It does not link FFmpeg development libraries; it loads the NVIDIA driver libraries (`libcuda.so.1`, `libnvidia-encode.so.1`) at runtime for NVENC (needs a driver with NVENC API 12.1, 530 or newer) and `libva.so.2`/`libva-drm.so.2` for VAAPI. When the capture and encoder sizes match, NVENC reads the XShm buffer directly, saving one host-to-device copy; VAAPI runs in process under the same feature (DRI3 DMA-BUF zero copy, or XShm upload without DRI3, then VPP to NV12 and VCN; H.264/HEVC). CPU encoding and `UU_CAPTURE=grab` keep the FFmpeg process; if the in-process hardware session fails it falls back to the matching FFmpeg process pipeline, then to the CPU. The default build and the release packages need neither feature.
### 3. Log in
@@ -429,12 +429,21 @@ sends about 12 kbit/s. `UU_VIDEO_CODEC=h265` selects H.265 explicitly for A/B di
`UU_VIDEO_ENCODER=vaapi` selects Intel/AMD H.264 VAAPI; `UU_VAAPI_DEVICE` defaults to
`/dev/dri/renderD128` (render-node permissions required). No automatic VAAPI selection.
Both FFmpeg process and `--features inproc-av` upload NV12 frames to a VAAPI frame pool.
The FFmpeg process uploads NV12 frames to a VAAPI frame pool; a self-built `--features inproc-av`
binary instead runs VAAPI in process on X11 (a DRI3 DMA-BUF of the root window zero-copy, or XShm
upload without DRI3, then VPP to NV12 and VCN; H.264/HEVC).
VBR uses the viewer bitrate preset and CQ-derived minimum QP, no B-frames, async depth 1,
AUD and the existing GOP/VFR timing. Unsupported device/driver/initialization falls back to libx264
with a warning. Accepted on Intel UHD 770 (iHD) with the default FFmpeg process pipeline and the official Mac
viewer (presets, fps, scaling, reconnect; no fallback, ~40% less CPU than libx264). AMD and the `inproc-av`
VAAPI path are not hardware-verified.
viewer (presets, fps, scaling, reconnect; no fallback, ~40% less CPU than libx264 — a figure from the
FFmpeg-process-pipeline era). AMD iGPUs are verified on
the **in-process engine** (`--features inproc-av`): end-to-end sessions with the official Mac viewer passed
acceptance on HawkPoint (Mesa 26.2.3) and Cezanne (Mesa 26.2.5) — HEVC, presets, fps, scaling; on Cezanne,
H.264 and both capture paths were additionally probe-verified: the XShm upload (decodable bitstream, correct
picture) and the DRI3 zero-copy import. Measured on one machine (1920x1200@60, Cezanne iGPU + RTX 3050 Ti):
in-process VAAPI costs about **1/12** of the per-frame encode CPU of libx264 (FFmpeg pipeline, superfast)
and about **1/5** of in-process NVENC. The FFmpeg process pipeline (what the release packages use) is not
separately verified on AMD.
The default VAAPI VBV buffer (100-500 kbit) squeezes keyframes hard at high resolution: on a 2560x1440
H.264 VAAPI stream capped at 30 Mbit/s the first IDR measured 59,910 bytes and the picture blurred at
@@ -490,8 +499,7 @@ Keep the unit name `uuremote-native.service` (not a template `[email protected]
Optional VAAPI: `intel-media-driver` (modern Intel) or `libva-mesa-driver` (AMD), `libva-utils`
for `vainfo`. Optional KMS: `libdrm`. Virtual Xfce: `xorg-server-xvfb xorg-xauth xorg-xdpyinfo
xfce4-session xfce4-panel xfdesktop xfwm4 gnome-terminal`.
Rolling Arch FFmpeg may be newer than supported `inproc-av` (FFmpeg 6.x); use the default process
pipeline or build against a separate FFmpeg 6 prefix. Do not downgrade the system FFmpeg.
If building `viewer-ffmpeg`, rolling Arch FFmpeg may be newer than supported 6.x; use a separate FFmpeg 6 prefix rather than downgrading system FFmpeg (`inproc-av` native NVENC on the host does not depend on FFmpeg).
User-bus headless setup and the official Mac checklist: [virtual desktop acceptance](contrib/virtual-desktop/ACCEPTANCE.md).
### Experimental KMS / 实验后端
+8 -8
View File
@@ -103,7 +103,7 @@ GNOME Portal 的控制通道不再等待授权窗口或第一帧画面:先通
| 功能 | 状态 |
|---|---|
| 多显示器切换 | Hyprland 原生 Wayland 后端实验性支持,见上文 |
| 真彩(4:4:4) | 支持(v0.1.15):NVENC / libx264 的默认 FFmpeg 管线输出 H.264 High 4:4:4;VAAPI 和 `inproc-av` 管线仍为 4:2:0,并如实回报给客户端 |
| 真彩(4:4:4) | 支持(v0.1.15):NVENC / libx264 的默认 FFmpeg 管线输出 H.264 High 4:4:4;`inproc-av` 原生 NVENC 输出 H.264 High 4:4:4 / HEVC FRExt;VAAPI 管线仍为 4:2:0,并如实回报给客户端 |
| HDR | 未实现 |
| 超级屏 / 虚拟显示器 | 未实现 |
| 被控端防窥模式 | 未实现 |
@@ -157,6 +157,8 @@ sudo apt install build-essential pkg-config \
sudo apt install libwayland-dev libxkbcommon-dev wl-clipboard
# 可选:GNOME/KDE portal 后端
sudo apt install libglib2.0-dev libpipewire-0.3-dev libei-dev
# 可选:inproc-av 的 VAAPI/DRI3 引擎编译所需(libva 本体运行时加载)
sudo apt install libva-dev libgbm-dev libx11-xcb-dev libxcb-dri3-dev libxcb-damage0-dev
```
- 需要 Rust 工具链(<https://rustup.rs>)。
@@ -171,9 +173,8 @@ cargo build --release --features wayland # + Hyprland
cargo build --release --features wayland,portal # + GNOME/KDE Wayland
```
可选的 `--features inproc-av`(进程内编码)和控制端的 `viewer-ffmpeg` 只支持 **FFmpeg 6.x**(`pkg-config --modversion libavcodec` 为 60.x)。
FFmpeg 7 及以上会在依赖 `ffmpeg-sys-next` 里报 `avfft.h file not found`——它的构建脚本先于本项目运行,所以无法提前给出更清楚的提示(#63)。
默认构建和预编译包都不需要这两个特性。
控制端的 `viewer-ffmpeg` 只支持 **FFmpeg 6.x**(`pkg-config --modversion libavcodec` 为 60.x);FFmpeg 7 及以上会在依赖 `ffmpeg-sys-next` 里报 `avfft.h file not found`(#63)。
服务端可选的 `--features inproc-av` 是 X11 下的进程内硬件编码:不链接 FFmpeg 开发库,NVENC 运行时加载 NVIDIA 驱动库(`libcuda.so.1`、`libnvidia-encode.so.1`,需要支持 NVENC API 12.1 的驱动,530 及以上),VAAPI 运行时加载 `libva.so.2`/`libva-drm.so.2`。截屏与编码尺寸一致时 NVENC 直接读取 XShm 缓冲,省去一次主机到显存的拷贝;VAAPI 在同一特性下也进程内运行(DRI3 DMA-BUF 零拷贝,无 DRI3 时 XShm 上传,VPP 转 NV12 后送 VCN,支持 H.264/H.265)。CPU 编码或 `UU_CAPTURE=grab` 仍走外部 FFmpeg 进程;进程内硬件编码启动失败时先回退到对应的 FFmpeg 进程管线,再回退到 CPU。默认构建和预编译包都不需要该特性。
### 3. 登录
@@ -384,9 +385,9 @@ HEVC 的访问单元按 RFC 7798 的图片边界切分,不依赖 AUD:`hevc_v
设置 `UU_VIDEO_ENCODER=vaapi` 使用 Intel/AMD H.264 VAAPI;`UU_VAAPI_DEVICE` 默认
`/dev/dri/renderD128`,运行用户需要 render 节点权限。不会自动选择 VAAPI;设备或驱动初始化
失败会明确警告并回退 libx264。外部 FFmpeg 和 `--features inproc-av` 均先转 NV12 再上传硬件帧。
失败会明确警告并回退 libx264。外部 FFmpeg 先转 NV12 再上传硬件帧。
VBR 使用控制端码率预设及 CQ 对应的最小 QP,无 B 帧、async depth=1,并保留现有 GOP/VFR 规则。
Intel 核显(UHD 770,iHD 驱动)已实机验收:默认的外部 FFmpeg 管线用官方 Mac 控制端测试了各档画质、帧率、缩放和重连,全程硬件编码、无回退,CPU 占用比 libx264 低约 40%。**AMD 显卡和 `--features inproc-av` 管线的 VAAPI 尚未实机验证。**
Intel 核显(UHD 770,iHD 驱动)已实机验收:默认的外部 FFmpeg 管线用官方 Mac 控制端测试了各档画质、帧率、缩放和重连,全程硬件编码、无回退,CPU 占用比 libx264 低约 40%(该数字为 FFmpeg 进程管线时代的测量)。AMD 核显已在**进程内原生引擎**上实机验收(`--features inproc-av`):官方 Mac 客户端的端到端会话在 HawkPoint(Mesa 26.2.3)与 Cezanne(Mesa 26.2.5)上通过验收(HEVC,含画质档、帧率、缩放);Cezanne 上另以探测验证了 H.264 与两条采集路径——XShm 上传(码流可解码、画面正确)与 DRI3 DMA-BUF 零拷贝导入。同机实测(1920×1200@60,Cezanne iGPU + RTX 3050 Ti 双显卡):进程内 VAAPI 每帧编码 CPU 约为 libx264(FFmpeg 管线 superfast)的 **1/12**、进程内 NVENC 的 **1/5**;外部 FFmpeg 进程管线(预编译包使用)在 AMD 上未单独验证。
高分辨率下 VAAPI 的默认 VBV 缓冲(100–500 kbit)会把关键帧压得过狠:2560×1440 H.264 VAAPI、30M
上限、静止画面上实测首个 IDR 只有 59,910 字节,画面每隔一个 GOP 糊一下再慢慢恢复。设
@@ -451,8 +452,7 @@ WantedBy=graphical-session.target
VAAPI 可选包:Intel 新显卡用 `intel-media-driver`,AMD 用 `libva-mesa-driver`,`libva-utils` 提供
`vainfo`。KMS 依赖 `libdrm`。虚拟 Xfce 依赖 `xorg-server-xvfb xorg-xauth xorg-xdpyinfo
xfce4-session xfce4-panel xfdesktop xfwm4 gnome-terminal`。
滚动 Arch 的 FFmpeg 可能超出 `inproc-av` 支持的 6.x;请使用默认进程方案或独立的 FFmpeg 6 前缀,
不要降级系统 FFmpeg。
若编译包含 `viewer-ffmpeg`,滚动 Arch 的 FFmpeg 可能超出其支持的 6.x;请使用独立的 FFmpeg 6 前缀,不要降级系统 FFmpeg(服务端的 `--features inproc-av` 原生 NVENC 不依赖 FFmpeg)。
### 实验性 KMS 截屏
+133
View File
@@ -0,0 +1,133 @@
//! Standalone benchmark tool measuring:
//! 1. Host Memory Copy Bandwidth & CPU Cost (What FFmpeg/Sway CPU swizzling does)
//! 2. NVENC H.265 Hardware Throughput & Latency (1920x1200 @ ultra-low-latency)
//! 3. NVENC reading a registered (pinned) host buffer
use std::time::Instant;
fn main() {
println!("=========================================================");
println!(" UU Remote Linux: Encoder Pipeline Benchmark Suite ");
println!("=========================================================");
println!();
let width: u32 = 1920;
let height: u32 = 1200;
let frames: usize = 300; // 5 seconds of 60fps video
let frame_bytes: usize = (width * height * 4) as usize;
println!("Test Configuration:");
println!(" Resolution : {}x{}", width, height);
println!(" Frame Size : {:.2} MB (ARGB32)", frame_bytes as f64 / 1024.0 / 1024.0);
println!(" Iterations : {} frames (Equivalent to 5.0s @ 60 FPS)", frames);
println!();
// Simulating XShm framebuffer
let mut host_pixels = vec![0x55u8; frame_bytes];
// Benchmark 1: CPU Memcpy (External FFmpeg stdin / Sway wlr-screencopy CPU copy baseline)
println!("[1] Baseline: CPU Frame Memcpy Overhead (IPC / Process Pipe Cost)");
let mut sink = vec![0u8; frame_bytes];
let start_memcpy = Instant::now();
for _ in 0..frames {
sink.copy_from_slice(&host_pixels);
std::hint::black_box(&sink);
}
let duration_memcpy = start_memcpy.elapsed();
let memcpy_per_frame = duration_memcpy.as_secs_f64() * 1000.0 / frames as f64;
let memcpy_gbps = (frame_bytes * frames) as f64 / duration_memcpy.as_secs_f64() / 1e9;
let cpu_load_60fps = (memcpy_per_frame / (1000.0 / 60.0)) * 100.0;
println!(" -> Total CPU copy time : {:.2?}", duration_memcpy);
println!(" -> Per-frame copy time : {:.3} ms", memcpy_per_frame);
println!(" -> Memory Bus Bandwidth : {:.2} GB/s", memcpy_gbps);
println!(" -> CPU Core Saturation : {:.1}% of 1 core consumed solely by memcpy at 60 FPS", cpu_load_60fps);
println!();
// Benchmark 2: in-process NVENC reading a registered host buffer (device copy when registration fails)
#[cfg(feature = "inproc-av")]
{
extern "C" {
fn nvenc_create(config: *const NvencConfigFFI) -> *mut std::ffi::c_void;
fn nvenc_register_host_buffer(
engine: *mut std::ffi::c_void,
host_ptr: *mut std::ffi::c_void,
size: usize,
) -> i32;
fn nvenc_encode_frame(
engine: *mut std::ffi::c_void,
input_argb: *const u8,
pitch: u32,
force_idr: i32,
out_data: *mut *const u8,
out_len: *mut u32,
) -> i32;
fn nvenc_destroy(engine: *mut std::ffi::c_void);
}
// Must match NvencConfig in c_engine/nvenc_engine.h.
#[repr(C)]
struct NvencConfigFFI {
width: u32,
height: u32,
fps: u32,
gop: u32,
cq: u32,
target_bps: i64,
max_bps: i64,
buffer_bits: i64,
is_hevc: i32,
chroma_444: i32,
}
let cfg = NvencConfigFFI {
width,
height,
fps: 60,
gop: 120,
cq: 18,
target_bps: 16_000_000,
max_bps: 30_000_000,
buffer_bits: 500_000,
is_hevc: 1, // H.265 / HEVC
chroma_444: i32::from(std::env::var("UU_BENCH_CHROMA").as_deref() == Ok("444")),
};
println!("[2] In-process NVENC H.265 encoding");
unsafe {
let engine = nvenc_create(&cfg);
if engine.is_null() {
println!(" [SKIP] NVENC driver not available on this system.");
} else {
let registered = nvenc_register_host_buffer(engine, host_pixels.as_mut_ptr() as *mut _, frame_bytes) == 0;
println!(" -> Input path : {}", if registered { "registered host buffer (DMA)" } else { "cuMemcpyHtoD copy" });
let input = if registered { std::ptr::null() } else { host_pixels.as_ptr() };
let mut out_data: *const u8 = std::ptr::null();
let mut out_len: u32 = 0;
// Warmup
let _ = nvenc_encode_frame(engine, input, width * 4, 1, &mut out_data, &mut out_len);
let start = Instant::now();
let (mut encoded_bytes, mut encoded_frames) = (0u64, 0usize);
for i in 0..frames {
if nvenc_encode_frame(engine, input, width * 4, i32::from(i == 0), &mut out_data, &mut out_len) == 0 {
encoded_bytes += u64::from(out_len);
encoded_frames += 1;
}
}
let elapsed = start.elapsed();
if encoded_frames == 0 {
println!(" [FAIL] no frame was encoded");
} else {
let per_frame = elapsed.as_secs_f64() * 1000.0 / encoded_frames as f64;
println!(" -> Encoded frames : {encoded_frames}/{frames} in {elapsed:.2?}");
println!(" -> Per-frame latency : {per_frame:.3} ms ({:.1} FPS max)", 1000.0 / per_frame);
println!(" -> Output bitrate : {:.2} Mbps average", encoded_bytes as f64 * 8.0 / elapsed.as_secs_f64() / 1e6);
if registered {
println!(" -> Host-to-device copy : none ({:.0} MB/s at 60 fps avoided)", frame_bytes as f64 * 60.0 / 1024.0 / 1024.0);
}
}
nvenc_destroy(engine);
}
}
}
}
+24 -4
View File
@@ -1,6 +1,10 @@
fn main() {
println!("cargo:rerun-if-changed=c_engine/portal_clipboard.inc");
println!("cargo:rerun-if-changed=c_engine/x11_engine.c");
println!("cargo:rerun-if-changed=c_engine/nvenc_engine.c");
println!("cargo:rerun-if-changed=c_engine/nvenc_engine.h");
println!("cargo:rerun-if-changed=vendor/ffnvcodec/ffnvcodec/dynlink_cuda.h");
println!("cargo:rerun-if-changed=vendor/ffnvcodec/ffnvcodec/nvEncodeAPI.h");
println!("cargo:rerun-if-changed=c_engine/clipboard.c");
println!("cargo:rerun-if-changed=c_engine/assist_window.c");
println!("cargo:rerun-if-changed=c_engine/viewer_window.c");
@@ -21,11 +25,27 @@ fn main() {
build.compile("uu_kms_engine");
println!("cargo:rustc-link-lib=drm");
}
cc::Build::new()
.file("c_engine/x11_engine.c")
let mut x11_build = cc::Build::new();
x11_build.file("c_engine/x11_engine.c")
.file("c_engine/clipboard.c")
.file("c_engine/pam_check.c")
.compile("x11_engine");
.file("c_engine/pam_check.c");
if std::env::var_os("CARGO_FEATURE_INPROC_AV").is_some() {
x11_build.file("c_engine/nvenc_engine.c");
x11_build.file("c_engine/vaapi_engine.c");
x11_build.file("c_engine/dri3_capture.c");
for f in ["c_engine/vaapi_engine.c", "c_engine/vaapi_engine.h", "c_engine/dri3_capture.c", "c_engine/dri3_capture.h"] {
println!("cargo:rerun-if-changed={f}");
}
x11_build.include("vendor/ffnvcodec");
// libva/libva-drm are dlopen'ed at runtime (like NVENC's libcuda); these are
// link-time-only deps of the DRI3/VAAPI engines and must stay out of default builds.
println!("cargo:rustc-link-lib=gbm");
println!("cargo:rustc-link-lib=X11-xcb");
println!("cargo:rustc-link-lib=xcb");
println!("cargo:rustc-link-lib=xcb-dri3");
println!("cargo:rustc-link-lib=xcb-damage");
}
x11_build.compile("x11_engine");
println!("cargo:rustc-link-lib=X11");
println!("cargo:rustc-link-lib=Xext");
+242
View File
@@ -0,0 +1,242 @@
#include "dri3_capture.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <fcntl.h>
#include <poll.h>
#include <X11/Xlib.h>
#include <X11/Xlib-xcb.h>
#include <xcb/xcb.h>
#include <xcb/dri3.h>
#include <xcb/damage.h>
#include <gbm.h>
struct Dri3Capture {
Display *dpy;
xcb_connection_t *conn;
xcb_screen_t *screen;
uint32_t width;
uint32_t height;
int drm_fd;
struct gbm_device *gbm;
struct gbm_bo *bo;
int bo_fd;
uint32_t stride;
xcb_pixmap_t pix;
xcb_gcontext_t gc;
xcb_damage_damage_t damage;
uint8_t damage_event_code;
};
Dri3Capture* dri3_capture_create(const char* display_name, uint32_t width, uint32_t height) {
Display *dpy = XOpenDisplay(display_name);
if (!dpy) {
fprintf(stderr, "[dri3_capture] failed to open display %s\n", display_name ? display_name : "default");
return NULL;
}
// Events are read with xcb_poll_for_event below, so XCB, not Xlib, must own the queue
XSetEventQueueOwner(dpy, XCBOwnsEventQueue);
xcb_connection_t *conn = XGetXCBConnection(dpy);
const xcb_setup_t *setup = xcb_get_setup(conn);
xcb_screen_t *screen = xcb_setup_roots_iterator(setup).data;
if (!screen) {
XCloseDisplay(dpy);
return NULL;
}
// Check the extension first: a DRI3 request on a server without it breaks the connection,
// and the Xlib cleanup below then exits the process.
const xcb_query_extension_reply_t *dri3_ext = xcb_get_extension_data(conn, &xcb_dri3_id);
if (!dri3_ext || !dri3_ext->present) {
fprintf(stderr, "[dri3_capture] DRI3 extension not supported on display\n");
XCloseDisplay(dpy);
return NULL;
}
// Query DRI3
xcb_dri3_query_version_cookie_t qvc = xcb_dri3_query_version(conn, 1, 2);
xcb_dri3_query_version_reply_t *qvr = xcb_dri3_query_version_reply(conn, qvc, NULL);
if (!qvr) {
fprintf(stderr, "[dri3_capture] DRI3 extension not supported on display\n");
XCloseDisplay(dpy);
return NULL;
}
free(qvr);
// Open DRM device via DRI3
xcb_dri3_open_cookie_t oc = xcb_dri3_open(conn, screen->root, 0);
xcb_dri3_open_reply_t *or = xcb_dri3_open_reply(conn, oc, NULL);
if (!or) {
fprintf(stderr, "[dri3_capture] xcb_dri3_open failed\n");
XCloseDisplay(dpy);
return NULL;
}
int drm_fd = xcb_dri3_open_reply_fds(conn, or)[0];
free(or);
uint32_t w = width > 0 ? width : screen->width_in_pixels;
uint32_t h = height > 0 ? height : screen->height_in_pixels;
struct gbm_device *gbm = gbm_create_device(drm_fd);
if (!gbm) {
fprintf(stderr, "[dri3_capture] gbm_create_device failed\n");
close(drm_fd);
XCloseDisplay(dpy);
return NULL;
}
struct gbm_bo *bo = gbm_bo_create(gbm, w, h, GBM_FORMAT_XRGB8888, GBM_BO_USE_RENDERING | GBM_BO_USE_LINEAR);
if (!bo) {
fprintf(stderr, "[dri3_capture] gbm_bo_create failed\n");
gbm_device_destroy(gbm);
close(drm_fd);
XCloseDisplay(dpy);
return NULL;
}
int bo_fd = gbm_bo_get_fd(bo);
uint32_t stride = gbm_bo_get_stride(bo);
xcb_pixmap_t pix = xcb_generate_id(conn);
xcb_generic_error_t *err = NULL;
// XCB takes ownership of the fd it sends and closes it after transmission; only a
// duplicate goes to X11, bo_fd stays ours for the VAAPI import and teardown.
int send_fd = fcntl(bo_fd, F_DUPFD_CLOEXEC, 3);
if (send_fd < 0) {
close(bo_fd);
gbm_bo_destroy(bo);
gbm_device_destroy(gbm);
close(drm_fd);
XCloseDisplay(dpy);
return NULL;
}
xcb_void_cookie_t pc = xcb_dri3_pixmap_from_buffer_checked(
conn, pix, screen->root, stride * h, w, h, stride, screen->root_depth, 32, send_fd
);
err = xcb_request_check(conn, pc);
if (err) {
fprintf(stderr, "[dri3_capture] xcb_dri3_pixmap_from_buffer failed: %d\n", err->error_code);
free(err);
close(bo_fd);
gbm_bo_destroy(bo);
gbm_device_destroy(gbm);
close(drm_fd);
XCloseDisplay(dpy);
return NULL;
}
xcb_gcontext_t gc = xcb_generate_id(conn);
// The default ClipByChildren copies only the root background (solid black); XShmGetImage on the root
// includes child windows, so the blit must too.
uint32_t gc_mode = XCB_SUBWINDOW_MODE_INCLUDE_INFERIORS;
xcb_create_gc(conn, gc, pix, XCB_GC_SUBWINDOW_MODE, &gc_mode);
// Setup XDamage on root window
const xcb_query_extension_reply_t *ext = xcb_get_extension_data(conn, &xcb_damage_id);
uint8_t damage_event_code = 0;
xcb_damage_damage_t damage = 0;
if (ext && ext->present) {
damage_event_code = ext->first_event;
xcb_damage_query_version(conn, 1, 1);
damage = xcb_generate_id(conn);
xcb_damage_create(conn, damage, screen->root, XCB_DAMAGE_REPORT_LEVEL_NON_EMPTY);
}
Dri3Capture* cap = (Dri3Capture*)calloc(1, sizeof(Dri3Capture));
cap->dpy = dpy;
cap->conn = conn;
cap->screen = screen;
cap->width = w;
cap->height = h;
cap->drm_fd = drm_fd;
cap->gbm = gbm;
cap->bo = bo;
cap->bo_fd = bo_fd;
cap->stride = stride;
cap->pix = pix;
cap->gc = gc;
cap->damage = damage;
cap->damage_event_code = damage_event_code;
fprintf(stderr, "[dri3_capture] Initialized: %ux%u (stride: %u, dma_fd: %d) Pure GPU Blit ready!\n",
w, h, stride, bo_fd);
return cap;
}
int dri3_capture_blit(Dri3Capture* cap) {
if (!cap) return -1;
xcb_copy_area(cap->conn, cap->screen->root, cap->pix, cap->gc, 0, 0, 0, 0, cap->width, cap->height);
// Round trip: the server has executed (and flushed to the GPU) the copy before VAAPI reads the BO;
// xcb_flush alone only sends the request.
xcb_get_input_focus_reply_t *r = xcb_get_input_focus_reply(cap->conn, xcb_get_input_focus(cap->conn), NULL);
if (!r) return -1;
free(r);
return 0;
}
int dri3_capture_wait_damage(Dri3Capture* cap, int timeout_ms) {
if (!cap) return -1;
// Check queued events
xcb_generic_event_t *ev = NULL;
int damaged = 0;
while ((ev = xcb_poll_for_event(cap->conn)) != NULL) {
if ((ev->response_type & 0x7f) == (cap->damage_event_code + XCB_DAMAGE_NOTIFY)) {
damaged = 1;
}
free(ev);
}
if (damaged) {
xcb_damage_subtract(cap->conn, cap->damage, XCB_NONE, XCB_NONE);
xcb_flush(cap->conn);
return 1;
}
int fd = xcb_get_file_descriptor(cap->conn);
struct pollfd pfd = { .fd = fd, .events = POLLIN, .revents = 0 };
int ret = poll(&pfd, 1, timeout_ms);
if (ret > 0 && (pfd.revents & POLLIN)) {
// Readable but not yet a whole event: polling again would return at once and spin a core
ev = xcb_poll_for_event(cap->conn);
if (!ev) ev = xcb_wait_for_event(cap->conn);
if (!ev) return -1;
do {
if ((ev->response_type & 0x7f) == (cap->damage_event_code + XCB_DAMAGE_NOTIFY)) {
damaged = 1;
}
free(ev);
} while ((ev = xcb_poll_for_event(cap->conn)) != NULL);
if (damaged) {
xcb_damage_subtract(cap->conn, cap->damage, XCB_NONE, XCB_NONE);
xcb_flush(cap->conn);
return 1;
}
}
return ret < 0 ? -1 : 0;
}
int dri3_capture_get_dma_fd(Dri3Capture* cap) {
return cap ? cap->bo_fd : -1;
}
uint32_t dri3_capture_get_stride(Dri3Capture* cap) {
return cap ? cap->stride : 0;
}
void dri3_capture_destroy(Dri3Capture* cap) {
if (!cap) return;
if (cap->damage) xcb_damage_destroy(cap->conn, cap->damage);
if (cap->gc) xcb_free_gc(cap->conn, cap->gc);
if (cap->pix) xcb_free_pixmap(cap->conn, cap->pix);
if (cap->bo_fd >= 0) close(cap->bo_fd);
if (cap->bo) gbm_bo_destroy(cap->bo);
if (cap->gbm) gbm_device_destroy(cap->gbm);
if (cap->drm_fd >= 0) close(cap->drm_fd);
if (cap->dpy) XCloseDisplay(cap->dpy);
free(cap);
}
+33
View File
@@ -0,0 +1,33 @@
#ifndef UU_DRI3_CAPTURE_H
#define UU_DRI3_CAPTURE_H
#include <stdint.h>
#include <stddef.h>
#ifdef __cplusplus
extern "C" {
#endif
typedef struct Dri3Capture Dri3Capture;
Dri3Capture* dri3_capture_create(const char* display_name, uint32_t width, uint32_t height);
// Trigger GPU blit from Root Window to the DRI3 DMA-BUF Pixmap.
// Returns 0 on success, negative on error.
int dri3_capture_blit(Dri3Capture* cap);
// Wait for next XDamage event or timeout (ms).
// Returns 1 if damaged, 0 if timeout, negative on error.
int dri3_capture_wait_damage(Dri3Capture* cap, int timeout_ms);
// Get the exported DMA-BUF file descriptor and stride
int dri3_capture_get_dma_fd(Dri3Capture* cap);
uint32_t dri3_capture_get_stride(Dri3Capture* cap);
void dri3_capture_destroy(Dri3Capture* cap);
#ifdef __cplusplus
}
#endif
#endif // UU_DRI3_CAPTURE_H
+580
View File
@@ -0,0 +1,580 @@
#include "nvenc_engine.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <dlfcn.h>
#include <ffnvcodec/dynlink_cuda.h>
#include <ffnvcodec/nvEncodeAPI.h>
// Struct revisions come from the vendored nvEncodeAPI.h (*_VER macros), so the session must be
// opened with that header's NVENCAPI_VERSION; older drivers are rejected in nvenc_create.
#define CU_MEMHOSTREGISTER_DEVICEMAP 0x02
typedef CUresult CUDAAPI (*tcuMemHostRegister)(void *p, size_t bytesize, unsigned int Flags);
typedef CUresult CUDAAPI (*tcuMemHostGetDevicePointer)(CUdeviceptr *pdptr, void *p, unsigned int Flags);
typedef CUresult CUDAAPI (*tcuMemHostUnregister)(void *p);
typedef NVENCSTATUS NVENCAPI (*tNvEncodeAPICreateInstance)(NV_ENCODE_API_FUNCTION_LIST *functionList);
typedef NVENCSTATUS NVENCAPI (*tNvEncodeAPIGetMaxSupportedVersion)(uint32_t *version);
struct NvencEngine {
void *cuda_lib;
void *nvenc_lib;
tcuInit *cuInit;
tcuDeviceGet *cuDeviceGet;
tcuDevicePrimaryCtxRetain *cuDevicePrimaryCtxRetain;
tcuDevicePrimaryCtxRelease *cuDevicePrimaryCtxRelease;
tcuCtxCreate_v2 *cuCtxCreate;
tcuCtxDestroy_v2 *cuCtxDestroy;
tcuCtxPushCurrent_v2 *cuCtxPushCurrent;
tcuCtxPopCurrent_v2 *cuCtxPopCurrent;
tcuMemAlloc_v2 *cuMemAlloc;
tcuMemFree_v2 *cuMemFree;
tcuMemcpyHtoD_v2 *cuMemcpyHtoD;
tcuMemHostRegister cuMemHostRegister;
tcuMemHostGetDevicePointer cuMemHostGetDevicePointer;
tcuMemHostUnregister cuMemHostUnregister;
CUdevice dev;
CUcontext cu_ctx;
int is_standalone_ctx; // cuCtxCreate fallback (primary context unavailable, seen on GB10)
CUdeviceptr cu_dptr;
size_t frame_bytes;
// Pinned XShm buffer registered for direct DMA reads
void *registered_host_ptr;
int is_zero_copy;
NV_ENCODE_API_FUNCTION_LIST fn;
void *encoder;
NV_ENC_REGISTER_RESOURCE reg_res;
NV_ENC_OUTPUT_PTR bitstream_buf;
uint32_t width;
uint32_t height;
int is_hevc;
uint8_t *output_staging;
size_t output_staging_cap;
};
// Returns 1 when the context was pushed; pop only in that case.
static int ctx_push(NvencEngine *engine) {
return engine->cuCtxPushCurrent && engine->cu_ctx && engine->cuCtxPushCurrent(engine->cu_ctx) == CUDA_SUCCESS;
}
static void ctx_pop(NvencEngine *engine, int pushed) {
if (pushed && engine->cuCtxPopCurrent) {
CUcontext dummy = NULL;
engine->cuCtxPopCurrent(&dummy);
}
}
NvencEngine* nvenc_create(const NvencConfig* config) {
if (!config || config->width == 0 || config->height == 0) {
return NULL;
}
NvencEngine *engine = (NvencEngine*)calloc(1, sizeof(NvencEngine));
if (!engine) return NULL;
engine->width = config->width;
engine->height = config->height;
engine->is_hevc = config->is_hevc;
// 1. Load CUDA driver
engine->cuda_lib = dlopen("libcuda.so.1", RTLD_NOW);
if (!engine->cuda_lib) {
fprintf(stderr, "[nvenc] dlopen libcuda.so.1 failed\n");
free(engine);
return NULL;
}
engine->cuInit = (tcuInit*)dlsym(engine->cuda_lib, "cuInit");
engine->cuDeviceGet = (tcuDeviceGet*)dlsym(engine->cuda_lib, "cuDeviceGet");
engine->cuDevicePrimaryCtxRetain = (tcuDevicePrimaryCtxRetain*)dlsym(engine->cuda_lib, "cuDevicePrimaryCtxRetain");
engine->cuDevicePrimaryCtxRelease = (tcuDevicePrimaryCtxRelease*)dlsym(engine->cuda_lib, "cuDevicePrimaryCtxRelease_v2");
if (!engine->cuDevicePrimaryCtxRelease)
engine->cuDevicePrimaryCtxRelease = (tcuDevicePrimaryCtxRelease*)dlsym(engine->cuda_lib, "cuDevicePrimaryCtxRelease");
engine->cuCtxCreate = (tcuCtxCreate_v2*)dlsym(engine->cuda_lib, "cuCtxCreate_v2");
engine->cuCtxDestroy = (tcuCtxDestroy_v2*)dlsym(engine->cuda_lib, "cuCtxDestroy_v2");
engine->cuCtxPushCurrent = (tcuCtxPushCurrent_v2*)dlsym(engine->cuda_lib, "cuCtxPushCurrent_v2");
engine->cuCtxPopCurrent = (tcuCtxPopCurrent_v2*)dlsym(engine->cuda_lib, "cuCtxPopCurrent_v2");
engine->cuMemAlloc = (tcuMemAlloc_v2*)dlsym(engine->cuda_lib, "cuMemAlloc_v2");
engine->cuMemFree = (tcuMemFree_v2*)dlsym(engine->cuda_lib, "cuMemFree_v2");
engine->cuMemcpyHtoD = (tcuMemcpyHtoD_v2*)dlsym(engine->cuda_lib, "cuMemcpyHtoD_v2");
engine->cuMemHostRegister = (tcuMemHostRegister)dlsym(engine->cuda_lib, "cuMemHostRegister_v2");
engine->cuMemHostGetDevicePointer = (tcuMemHostGetDevicePointer)dlsym(engine->cuda_lib, "cuMemHostGetDevicePointer_v2");
engine->cuMemHostUnregister = (tcuMemHostUnregister)dlsym(engine->cuda_lib, "cuMemHostUnregister");
if (!engine->cuInit || !engine->cuDeviceGet || !engine->cuDevicePrimaryCtxRetain ||
!engine->cuDevicePrimaryCtxRelease || !engine->cuCtxPushCurrent || !engine->cuCtxPopCurrent ||
!engine->cuMemAlloc || !engine->cuMemFree || !engine->cuMemcpyHtoD) {
fprintf(stderr, "[nvenc] Missing CUDA driver symbols\n");
nvenc_destroy(engine);
return NULL;
}
CUresult cures = engine->cuInit(0);
if (cures != CUDA_SUCCESS) {
fprintf(stderr, "[nvenc] cuInit failed: %d\n", cures);
nvenc_destroy(engine);
return NULL;
}
cures = engine->cuDeviceGet(&engine->dev, 0);
if (cures != CUDA_SUCCESS) {
fprintf(stderr, "[nvenc] cuDeviceGet failed: %d\n", cures);
nvenc_destroy(engine);
return NULL;
}
cures = engine->cuDevicePrimaryCtxRetain(&engine->cu_ctx, engine->dev);
if ((cures != CUDA_SUCCESS || !engine->cu_ctx) && engine->cuCtxCreate && engine->cuCtxDestroy) {
engine->cu_ctx = NULL;
// cuCtxCreate makes the new context current; the push below makes it current again
cures = engine->cuCtxCreate(&engine->cu_ctx, 0, engine->dev);
if (cures == CUDA_SUCCESS && engine->cu_ctx) {
CUcontext dummy = NULL;
engine->cuCtxPopCurrent(&dummy);
engine->is_standalone_ctx = 1;
}
}
if (cures != CUDA_SUCCESS || !engine->cu_ctx) {
fprintf(stderr, "[nvenc] CUDA context creation failed: %d\n", cures);
engine->cu_ctx = NULL;
nvenc_destroy(engine);
return NULL;
}
int pushed = ctx_push(engine);
if (!pushed) {
fprintf(stderr, "[nvenc] cuCtxPushCurrent failed\n");
nvenc_destroy(engine);
return NULL;
}
// 2. Load NVENC
engine->nvenc_lib = dlopen("libnvidia-encode.so.1", RTLD_NOW);
if (!engine->nvenc_lib) {
fprintf(stderr, "[nvenc] dlopen libnvidia-encode.so.1 failed\n");
goto fail;
}
// Like FFmpeg: the driver must support the header's API version, otherwise fall back.
tNvEncodeAPIGetMaxSupportedVersion get_max_ver = (tNvEncodeAPIGetMaxSupportedVersion)dlsym(
engine->nvenc_lib, "NvEncodeAPIGetMaxSupportedVersion");
uint32_t max_supported_ver = 0;
if (!get_max_ver || get_max_ver(&max_supported_ver) != NV_ENC_SUCCESS) {
fprintf(stderr, "[nvenc] NvEncodeAPIGetMaxSupportedVersion failed\n");
goto fail;
}
if (max_supported_ver < ((NVENCAPI_MAJOR_VERSION << 4) | NVENCAPI_MINOR_VERSION)) {
fprintf(stderr, "[nvenc] Driver NVENC API %u.%u is older than required %d.%d; update the NVIDIA driver\n",
max_supported_ver >> 4, max_supported_ver & 0xf, NVENCAPI_MAJOR_VERSION, NVENCAPI_MINOR_VERSION);
goto fail;
}
tNvEncodeAPICreateInstance create_inst = (tNvEncodeAPICreateInstance)dlsym(engine->nvenc_lib, "NvEncodeAPICreateInstance");
if (!create_inst) {
fprintf(stderr, "[nvenc] dlsym NvEncodeAPICreateInstance failed\n");
goto fail;
}
memset(&engine->fn, 0, sizeof(engine->fn));
engine->fn.version = NV_ENCODE_API_FUNCTION_LIST_VER;
NVENCSTATUS nvres = create_inst(&engine->fn);
if (nvres != NV_ENC_SUCCESS) {
fprintf(stderr, "[nvenc] NvEncodeAPICreateInstance failed: %d\n", nvres);
goto fail;
}
// 3. Open Encode Session
NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS open_params;
memset(&open_params, 0, sizeof(open_params));
open_params.version = NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER;
open_params.deviceType = NV_ENC_DEVICE_TYPE_CUDA;
open_params.device = engine->cu_ctx;
open_params.apiVersion = NVENCAPI_VERSION;
nvres = engine->fn.nvEncOpenEncodeSessionEx(&open_params, &engine->encoder);
if (nvres != NV_ENC_SUCCESS || !engine->encoder) {
fprintf(stderr, "[nvenc] nvEncOpenEncodeSessionEx failed: %d\n", nvres);
engine->encoder = NULL;
goto fail;
}
// 4. Query Preset Config
GUID codec_guid = engine->is_hevc ? NV_ENC_CODEC_HEVC_GUID : NV_ENC_CODEC_H264_GUID;
NV_ENC_PRESET_CONFIG preset_cfg;
memset(&preset_cfg, 0, sizeof(preset_cfg));
preset_cfg.version = NV_ENC_PRESET_CONFIG_VER;
preset_cfg.presetCfg.version = NV_ENC_CONFIG_VER;
nvres = engine->fn.nvEncGetEncodePresetConfigEx(
engine->encoder,
codec_guid,
NV_ENC_PRESET_P4_GUID,
NV_ENC_TUNING_INFO_ULTRA_LOW_LATENCY,
&preset_cfg
);
if (nvres != NV_ENC_SUCCESS) {
fprintf(stderr, "[nvenc] nvEncGetEncodePresetConfigEx failed: %d\n", nvres);
goto fail;
}
NV_ENC_CONFIG enc_cfg;
memcpy(&enc_cfg, &preset_cfg.presetCfg, sizeof(NV_ENC_CONFIG));
enc_cfg.gopLength = config->gop > 0 ? config->gop : NVENC_INFINITE_GOPLENGTH;
enc_cfg.frameIntervalP = 1; // Strict low-latency: zero B-frames
enc_cfg.rcParams.rateControlMode = NV_ENC_PARAMS_RC_VBR;
enc_cfg.rcParams.averageBitRate = (uint32_t)(config->target_bps > 0 ? config->target_bps : 10000000);
enc_cfg.rcParams.maxBitRate = (uint32_t)(config->max_bps > 0 ? config->max_bps : 15000000);
enc_cfg.rcParams.vbvBufferSize = (uint32_t)(config->buffer_bits > 0 ? config->buffer_bits : (config->max_bps > 0 ? config->max_bps / 10 : 1500000));
enc_cfg.rcParams.targetQuality = (uint8_t)(config->cq > 0 ? config->cq : 28);
NV_ENC_CAPS_PARAM caps_param;
memset(&caps_param, 0, sizeof(caps_param));
caps_param.version = NV_ENC_CAPS_PARAM_VER;
int caps_val = 0;
if (config->chroma_444) {
// GPUs without 4:4:4 encode fail here; the caller falls back to FFmpeg/libx264, which keep 4:4:4.
caps_param.capsToQuery = NV_ENC_CAPS_SUPPORT_YUV444_ENCODE;
if (engine->fn.nvEncGetEncodeCaps(engine->encoder, codec_guid, &caps_param, &caps_val) != NV_ENC_SUCCESS || caps_val <= 0) {
fprintf(stderr, "[nvenc] 4:4:4 encoding not supported for this codec\n");
goto fail;
}
// NVENC rejects AQ / temporal AQ in 4:4:4 mode.
enc_cfg.rcParams.enableAQ = 0;
enc_cfg.rcParams.enableTemporalAQ = 0;
} else {
enc_cfg.rcParams.enableAQ = 1;
// Temporal AQ only where the hardware/codec reports it
caps_param.capsToQuery = NV_ENC_CAPS_SUPPORT_TEMPORAL_AQ;
caps_val = 0;
enc_cfg.rcParams.enableTemporalAQ =
engine->fn.nvEncGetEncodeCaps(engine->encoder, codec_guid, &caps_param, &caps_val) == NV_ENC_SUCCESS && caps_val > 0;
}
if (engine->is_hevc) {
enc_cfg.profileGUID = config->chroma_444 ? NV_ENC_HEVC_PROFILE_FREXT_GUID : NV_ENC_HEVC_PROFILE_MAIN_GUID;
enc_cfg.encodeCodecConfig.hevcConfig.chromaFormatIDC = config->chroma_444 ? 3 : 1;
enc_cfg.encodeCodecConfig.hevcConfig.idrPeriod = enc_cfg.gopLength;
enc_cfg.encodeCodecConfig.hevcConfig.repeatSPSPPS = 1;
enc_cfg.encodeCodecConfig.hevcConfig.outputAUD = 1;
} else {
enc_cfg.profileGUID = config->chroma_444 ? NV_ENC_H264_PROFILE_HIGH_444_GUID : NV_ENC_H264_PROFILE_BASELINE_GUID;
enc_cfg.encodeCodecConfig.h264Config.chromaFormatIDC = config->chroma_444 ? 3 : 1;
enc_cfg.encodeCodecConfig.h264Config.idrPeriod = enc_cfg.gopLength;
enc_cfg.encodeCodecConfig.h264Config.repeatSPSPPS = 1;
enc_cfg.encodeCodecConfig.h264Config.outputAUD = 1;
}
// 5. Initialize Encoder
NV_ENC_INITIALIZE_PARAMS init_params;
memset(&init_params, 0, sizeof(init_params));
init_params.version = NV_ENC_INITIALIZE_PARAMS_VER;
init_params.encodeGUID = codec_guid;
init_params.presetGUID = NV_ENC_PRESET_P4_GUID;
init_params.encodeWidth = engine->width;
init_params.encodeHeight = engine->height;
init_params.darWidth = engine->width;
init_params.darHeight = engine->height;
init_params.frameRateNum = config->fps > 0 ? config->fps : 60;
init_params.frameRateDen = 1;
init_params.enablePTD = 1;
init_params.tuningInfo = NV_ENC_TUNING_INFO_ULTRA_LOW_LATENCY;
init_params.encodeConfig = &enc_cfg;
nvres = engine->fn.nvEncInitializeEncoder(engine->encoder, &init_params);
if (nvres != NV_ENC_SUCCESS) {
fprintf(stderr, "[nvenc] nvEncInitializeEncoder failed: %d\n", nvres);
goto fail;
}
// 6. Device input buffer (ARGB32), used until a host buffer is registered
engine->frame_bytes = (size_t)engine->width * engine->height * 4;
cures = engine->cuMemAlloc(&engine->cu_dptr, engine->frame_bytes);
if (cures != CUDA_SUCCESS) {
fprintf(stderr, "[nvenc] cuMemAlloc failed: %d\n", cures);
engine->cu_dptr = 0;
goto fail;
}
memset(&engine->reg_res, 0, sizeof(engine->reg_res));
engine->reg_res.version = NV_ENC_REGISTER_RESOURCE_VER;
engine->reg_res.resourceType = NV_ENC_INPUT_RESOURCE_TYPE_CUDADEVICEPTR;
engine->reg_res.width = engine->width;
engine->reg_res.height = engine->height;
engine->reg_res.pitch = engine->width * 4;
engine->reg_res.resourceToRegister = (void*)engine->cu_dptr;
engine->reg_res.bufferFormat = NV_ENC_BUFFER_FORMAT_ARGB;
nvres = engine->fn.nvEncRegisterResource(engine->encoder, &engine->reg_res);
if (nvres != NV_ENC_SUCCESS) {
fprintf(stderr, "[nvenc] nvEncRegisterResource failed: %d\n", nvres);
engine->reg_res.registeredResource = NULL;
goto fail;
}
// 7. Create bitstream output buffer
NV_ENC_CREATE_BITSTREAM_BUFFER bs_buf;
memset(&bs_buf, 0, sizeof(bs_buf));
bs_buf.version = NV_ENC_CREATE_BITSTREAM_BUFFER_VER;
nvres = engine->fn.nvEncCreateBitstreamBuffer(engine->encoder, &bs_buf);
if (nvres != NV_ENC_SUCCESS) {
fprintf(stderr, "[nvenc] nvEncCreateBitstreamBuffer failed: %d\n", nvres);
goto fail;
}
engine->bitstream_buf = bs_buf.bitstreamBuffer;
engine->output_staging_cap = 2 * 1024 * 1024;
engine->output_staging = (uint8_t*)malloc(engine->output_staging_cap);
if (!engine->output_staging) {
fprintf(stderr, "[nvenc] output buffer allocation failed\n");
goto fail;
}
ctx_pop(engine, pushed);
return engine;
fail:
// Pop the creation push first; nvenc_destroy pushes and pops on its own.
ctx_pop(engine, pushed);
nvenc_destroy(engine);
return NULL;
}
int nvenc_register_host_buffer(NvencEngine* engine, void* host_ptr, size_t size) {
if (!engine || !host_ptr || size < engine->frame_bytes) {
return -1;
}
if (!engine->cuMemHostRegister || !engine->cuMemHostGetDevicePointer || !engine->cuMemHostUnregister) {
return -2;
}
int pushed = ctx_push(engine);
if (!pushed) return -2;
// If already registered, unregister first
nvenc_unregister_host_buffer(engine);
CUresult cures = engine->cuMemHostRegister(host_ptr, size, CU_MEMHOSTREGISTER_DEVICEMAP);
if (cures != CUDA_SUCCESS) {
fprintf(stderr, "[nvenc] cuMemHostRegister failed: %d\n", cures);
ctx_pop(engine, pushed);
return -3;
}
CUdeviceptr dev_ptr = 0;
cures = engine->cuMemHostGetDevicePointer(&dev_ptr, host_ptr, 0);
if (cures != CUDA_SUCCESS || dev_ptr == 0) {
fprintf(stderr, "[nvenc] cuMemHostGetDevicePointer failed: %d\n", cures);
engine->cuMemHostUnregister(host_ptr);
ctx_pop(engine, pushed);
return -4;
}
// Register the pinned buffer before releasing the device buffer, so a failure keeps the copy path working
NV_ENC_REGISTER_RESOURCE new_reg;
memset(&new_reg, 0, sizeof(new_reg));
new_reg.version = NV_ENC_REGISTER_RESOURCE_VER;
new_reg.resourceType = NV_ENC_INPUT_RESOURCE_TYPE_CUDADEVICEPTR;
new_reg.width = engine->width;
new_reg.height = engine->height;
new_reg.pitch = engine->width * 4;
new_reg.resourceToRegister = (void*)dev_ptr;
new_reg.bufferFormat = NV_ENC_BUFFER_FORMAT_ARGB;
NVENCSTATUS nvres = engine->fn.nvEncRegisterResource(engine->encoder, &new_reg);
if (nvres != NV_ENC_SUCCESS) {
fprintf(stderr, "[nvenc] nvEncRegisterResource (pinned host) failed: %d\n", nvres);
engine->cuMemHostUnregister(host_ptr);
ctx_pop(engine, pushed);
return -5;
}
if (engine->reg_res.registeredResource) {
engine->fn.nvEncUnregisterResource(engine->encoder, engine->reg_res.registeredResource);
}
if (engine->cu_dptr) {
engine->cuMemFree(engine->cu_dptr);
engine->cu_dptr = 0;
}
engine->reg_res = new_reg;
engine->registered_host_ptr = host_ptr;
engine->is_zero_copy = 1;
ctx_pop(engine, pushed);
return 0;
}
// Caller has the context current.
void nvenc_unregister_host_buffer(NvencEngine* engine) {
if (!engine || !engine->is_zero_copy) return;
if (engine->reg_res.registeredResource && engine->encoder) {
engine->fn.nvEncUnregisterResource(engine->encoder, engine->reg_res.registeredResource);
memset(&engine->reg_res, 0, sizeof(engine->reg_res));
}
if (engine->registered_host_ptr && engine->cuMemHostUnregister) {
engine->cuMemHostUnregister(engine->registered_host_ptr);
engine->registered_host_ptr = NULL;
}
engine->is_zero_copy = 0;
}
int nvenc_encode_frame(
NvencEngine* engine,
const uint8_t* input_argb,
uint32_t pitch,
int force_idr,
const uint8_t** out_data,
uint32_t* out_len
) {
if (!engine || !out_data || !out_len || !engine->reg_res.registeredResource) {
return -1;
}
*out_data = NULL;
*out_len = 0;
int pushed = ctx_push(engine);
if (!pushed) return -1;
// Without a registered host buffer, copy the frame into the device buffer
if (!engine->is_zero_copy) {
if (!input_argb) {
ctx_pop(engine, pushed);
return -1;
}
CUresult cures = CUDA_SUCCESS;
if (pitch == engine->width * 4) {
cures = engine->cuMemcpyHtoD(engine->cu_dptr, input_argb, engine->frame_bytes);
} else {
for (uint32_t y = 0; y < engine->height && cures == CUDA_SUCCESS; y++) {
cures = engine->cuMemcpyHtoD(
engine->cu_dptr + (size_t)y * engine->width * 4,
input_argb + (size_t)y * pitch,
engine->width * 4
);
}
}
if (cures != CUDA_SUCCESS) {
fprintf(stderr, "[nvenc] cuMemcpyHtoD failed: %d\n", cures);
ctx_pop(engine, pushed);
return -2;
}
}
NV_ENC_MAP_INPUT_RESOURCE map_res;
memset(&map_res, 0, sizeof(map_res));
map_res.version = NV_ENC_MAP_INPUT_RESOURCE_VER;
map_res.registeredResource = engine->reg_res.registeredResource;
NVENCSTATUS nvres = engine->fn.nvEncMapInputResource(engine->encoder, &map_res);
if (nvres != NV_ENC_SUCCESS) {
fprintf(stderr, "[nvenc] nvEncMapInputResource failed: %d\n", nvres);
ctx_pop(engine, pushed);
return -3;
}
NV_ENC_PIC_PARAMS pic_params;
memset(&pic_params, 0, sizeof(pic_params));
pic_params.version = NV_ENC_PIC_PARAMS_VER;
pic_params.inputWidth = engine->width;
pic_params.inputHeight = engine->height;
pic_params.inputPitch = engine->width * 4;
pic_params.inputBuffer = map_res.mappedResource;
pic_params.outputBitstream = engine->bitstream_buf;
pic_params.bufferFmt = NV_ENC_BUFFER_FORMAT_ARGB;
pic_params.pictureStruct = NV_ENC_PIC_STRUCT_FRAME;
if (force_idr) {
pic_params.encodePicFlags = NV_ENC_PIC_FLAG_FORCEIDR;
}
nvres = engine->fn.nvEncEncodePicture(engine->encoder, &pic_params);
if (nvres != NV_ENC_SUCCESS) {
fprintf(stderr, "[nvenc] nvEncEncodePicture failed: %d\n", nvres);
engine->fn.nvEncUnmapInputResource(engine->encoder, map_res.mappedResource);
ctx_pop(engine, pushed);
return -4;
}
NV_ENC_LOCK_BITSTREAM lock_bs;
memset(&lock_bs, 0, sizeof(lock_bs));
lock_bs.version = NV_ENC_LOCK_BITSTREAM_VER;
lock_bs.outputBitstream = engine->bitstream_buf;
lock_bs.doNotWait = 0;
nvres = engine->fn.nvEncLockBitstream(engine->encoder, &lock_bs);
if (nvres != NV_ENC_SUCCESS) {
fprintf(stderr, "[nvenc] nvEncLockBitstream failed: %d\n", nvres);
engine->fn.nvEncUnmapInputResource(engine->encoder, map_res.mappedResource);
ctx_pop(engine, pushed);
return -5;
}
int rc = 0;
uint32_t size = lock_bs.bitstreamSizeInBytes;
if (size > engine->output_staging_cap) {
size_t cap = (size_t)size + 64 * 1024;
uint8_t *grown = (uint8_t*)realloc(engine->output_staging, cap);
if (grown) {
engine->output_staging = grown;
engine->output_staging_cap = cap;
} else {
fprintf(stderr, "[nvenc] output buffer allocation failed: %u bytes\n", size);
rc = -6;
}
}
if (rc == 0 && size > 0 && lock_bs.bitstreamBufferPtr) {
memcpy(engine->output_staging, lock_bs.bitstreamBufferPtr, size);
*out_data = engine->output_staging;
*out_len = size;
}
engine->fn.nvEncUnlockBitstream(engine->encoder, engine->bitstream_buf);
// The SDK allows unmapping only after the bitstream lock returned for this picture
engine->fn.nvEncUnmapInputResource(engine->encoder, map_res.mappedResource);
ctx_pop(engine, pushed);
return rc;
}
void nvenc_destroy(NvencEngine* engine) {
if (!engine) return;
int pushed = ctx_push(engine);
// Without a current context the NVENC/CUDA objects cannot be released safely; leak them.
if (pushed) {
if (engine->bitstream_buf && engine->encoder) {
engine->fn.nvEncDestroyBitstreamBuffer(engine->encoder, engine->bitstream_buf);
engine->bitstream_buf = NULL;
}
nvenc_unregister_host_buffer(engine);
if (engine->reg_res.registeredResource && engine->encoder) {
engine->fn.nvEncUnregisterResource(engine->encoder, engine->reg_res.registeredResource);
memset(&engine->reg_res, 0, sizeof(engine->reg_res));
}
if (engine->cu_dptr) {
engine->cuMemFree(engine->cu_dptr);
engine->cu_dptr = 0;
}
if (engine->encoder) {
engine->fn.nvEncDestroyEncoder(engine->encoder);
engine->encoder = NULL;
}
ctx_pop(engine, pushed);
}
if (engine->cu_ctx) {
if (engine->is_standalone_ctx) engine->cuCtxDestroy(engine->cu_ctx);
else engine->cuDevicePrimaryCtxRelease(engine->dev);
engine->cu_ctx = NULL;
}
free(engine->output_staging);
// libcuda/libnvidia-encode stay loaded: their driver threads outlive the session, and
// dlclose crashed them on the next encoder start (SEGV_MAPERR in libnvcuvid).
free(engine);
}
+59
View File
@@ -0,0 +1,59 @@
#ifndef UU_NVENC_ENGINE_H
#define UU_NVENC_ENGINE_H
#include <stdint.h>
#include <stddef.h>
#ifdef __cplusplus
extern "C" {
#endif
typedef struct NvencEngine NvencEngine;
typedef struct {
uint32_t width;
uint32_t height;
uint32_t fps;
uint32_t gop;
uint32_t cq;
int64_t target_bps;
int64_t max_bps;
int64_t buffer_bits;
int is_hevc; // 1 = H.265, 0 = H.264
int chroma_444; // 1 = 4:4:4 (H.264 High 4:4:4 / HEVC FRExt), 0 = 4:2:0
} NvencConfig;
// Initialize and open an NVENC encoding session. Returns NULL on error.
NvencEngine* nvenc_create(const NvencConfig* config);
// Register host shared memory (XShm buffer) as CUDA pinned memory that NVENC reads by DMA.
// Returns 0 on success, negative value on error.
int nvenc_register_host_buffer(NvencEngine* engine, void* host_ptr, size_t size);
// Unregister a previously registered host buffer (the CUDA context must be current).
void nvenc_unregister_host_buffer(NvencEngine* engine);
// Encode one frame.
// If host buffer was registered via nvenc_register_host_buffer, input_argb can be NULL
// and the encoder reads the registered host memory directly (no host-to-device copy).
// force_idr: 1 to force keyframe/IDR, 0 otherwise
// out_data: pointer to receive the bitstream buffer (valid until next encode call or destroy)
// out_len: pointer to receive the bitstream size in bytes
// Returns 0 on success, negative value on error.
int nvenc_encode_frame(
NvencEngine* engine,
const uint8_t* input_argb,
uint32_t pitch,
int force_idr,
const uint8_t** out_data,
uint32_t* out_len
);
// Destroy the NVENC engine and free all CUDA/NVENC resources.
void nvenc_destroy(NvencEngine* engine);
#ifdef __cplusplus
}
#endif
#endif // UU_NVENC_ENGINE_H
+835
View File
@@ -0,0 +1,835 @@
#include "vaapi_engine.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <fcntl.h>
#include <dlfcn.h>
#include <va/va.h>
#include <va/va_drm.h>
#include <va/va_drmcommon.h>
#include <va/va_vpp.h>
#include <va/va_enc_h264.h>
// Runtime-loaded libva, like the NVENC engine's libcuda/libnvidia-encode: an inproc-av build
// must not add a hard DT_NEEDED on libva for users who never touch VAAPI.
// Signatures come from the installed headers via __typeof__ so they cannot drift.
static struct {
void *libva, *libva_drm;
__typeof__(vaGetDisplayDRM) *vaGetDisplayDRM;
__typeof__(vaInitialize) *vaInitialize;
__typeof__(vaTerminate) *vaTerminate;
__typeof__(vaErrorStr) *vaErrorStr;
__typeof__(vaCreateSurfaces) *vaCreateSurfaces;
__typeof__(vaDestroySurfaces) *vaDestroySurfaces;
__typeof__(vaCreateConfig) *vaCreateConfig;
__typeof__(vaDestroyConfig) *vaDestroyConfig;
__typeof__(vaCreateContext) *vaCreateContext;
__typeof__(vaDestroyContext) *vaDestroyContext;
__typeof__(vaCreateBuffer) *vaCreateBuffer;
__typeof__(vaDestroyBuffer) *vaDestroyBuffer;
__typeof__(vaMapBuffer) *vaMapBuffer;
__typeof__(vaUnmapBuffer) *vaUnmapBuffer;
__typeof__(vaCreateImage) *vaCreateImage;
__typeof__(vaDestroyImage) *vaDestroyImage;
__typeof__(vaPutImage) *vaPutImage;
__typeof__(vaBeginPicture) *vaBeginPicture;
__typeof__(vaRenderPicture) *vaRenderPicture;
__typeof__(vaEndPicture) *vaEndPicture;
__typeof__(vaSyncSurface) *vaSyncSurface;
} RT;
static int vaapi_load_runtime(void) {
if (RT.libva) return 0;
RT.libva = dlopen("libva.so.2", RTLD_NOW | RTLD_GLOBAL);
RT.libva_drm = dlopen("libva-drm.so.2", RTLD_NOW);
if (!RT.libva || !RT.libva_drm) {
fprintf(stderr, "[vaapi] dlopen libva/libva-drm failed\n");
return -1;
}
// libva-drm depends on libva, so one lookup covers both libraries
#define VA_LOAD(sym) do { \
*(void**)&RT.sym = dlsym(RT.libva_drm, #sym) ? dlsym(RT.libva_drm, #sym) : dlsym(RT.libva, #sym); \
if (!RT.sym) { dlclose(RT.libva_drm); dlclose(RT.libva); RT.libva = RT.libva_drm = NULL; return -1; } \
} while (0)
VA_LOAD(vaGetDisplayDRM);
VA_LOAD(vaInitialize);
VA_LOAD(vaTerminate);
VA_LOAD(vaErrorStr);
VA_LOAD(vaCreateSurfaces);
VA_LOAD(vaDestroySurfaces);
VA_LOAD(vaCreateConfig);
VA_LOAD(vaDestroyConfig);
VA_LOAD(vaCreateContext);
VA_LOAD(vaDestroyContext);
VA_LOAD(vaCreateBuffer);
VA_LOAD(vaDestroyBuffer);
VA_LOAD(vaMapBuffer);
VA_LOAD(vaUnmapBuffer);
VA_LOAD(vaCreateImage);
VA_LOAD(vaDestroyImage);
VA_LOAD(vaPutImage);
VA_LOAD(vaBeginPicture);
VA_LOAD(vaRenderPicture);
VA_LOAD(vaEndPicture);
VA_LOAD(vaSyncSurface);
#undef VA_LOAD
return 0;
}
struct VaapiEngine {
int drm_fd;
VADisplay va_dpy;
uint32_t cap_width; // capture (RGB input) size
uint32_t cap_height;
uint32_t width; // encode (output) size
uint32_t height;
uint32_t fps;
uint32_t gop;
int64_t target_bps;
int64_t max_bps;
int64_t buffer_bits;
int is_hevc;
uint32_t frame_count;
// frames since the last IDR (periodic GOP), the alternating idr_pic_id, and an IDR owed after a lost frame
uint32_t since_idr;
uint32_t idr_pic_id;
int idr_pending;
// frame_num / POC bookkeeping since the last IDR; H264 state the driver does not derive itself
uint32_t frame_num;
uint32_t rc_mode;
// DMA-BUF backing
int dma_fd;
uint32_t stride;
// Surfaces & Contexts
VASurfaceID in_rgb_surf;
VASurfaceID nv12_surf;
VASurfaceID recon_surfs[2];
VAConfigID vpp_cfg;
VAContextID vpp_ctx;
VABufferID vpp_buf;
VARectangle src_rect, dst_rect;
VAConfigID enc_cfg;
VAContextID enc_ctx;
VABufferID coded_buf;
uint8_t *output_staging;
uint32_t output_staging_cap;
// Host upload path: a VAImage the CPU writes, then vaPutImage into in_rgb_surf
VAImage upload_img;
int upload_ready;
};
// The VPP pipeline buffer points at the engine-owned rects, so VPP scales cap -> output
static int vaapi_make_vpp_buffer(VaapiEngine* engine) {
VAProcPipelineParameterBuffer vpp_param;
memset(&vpp_param, 0, sizeof(vpp_param));
vpp_param.surface = engine->in_rgb_surf;
vpp_param.surface_region = &engine->src_rect;
vpp_param.output_region = &engine->dst_rect;
return RT.vaCreateBuffer(engine->va_dpy, engine->vpp_ctx, VAProcPipelineParameterBufferType,
sizeof(vpp_param), 1, &vpp_param, &engine->vpp_buf) == VA_STATUS_SUCCESS ? 0 : -1;
}
VaapiEngine* vaapi_create(const VaapiConfig* config) {
if (!config || config->width == 0 || config->height == 0 || config->cap_width == 0 || config->cap_height == 0) return NULL;
if (vaapi_load_runtime() != 0) return NULL;
VaapiEngine* engine = (VaapiEngine*)calloc(1, sizeof(VaapiEngine));
if (!engine) return NULL;
engine->cap_width = config->cap_width;
engine->cap_height = config->cap_height;
engine->width = config->width;
engine->height = config->height;
engine->fps = config->fps > 0 ? config->fps : 60;
engine->gop = config->gop > 0 ? config->gop : 30;
engine->target_bps = config->target_bps > 0 ? config->target_bps : 6000000;
engine->max_bps = config->max_bps > engine->target_bps ? config->max_bps : engine->target_bps;
engine->buffer_bits = config->buffer_bits > 0 ? config->buffer_bits : engine->max_bps;
engine->is_hevc = config->is_hevc;
engine->dma_fd = -1;
// calloc leaves surface/context IDs at 0, which is a valid VA ID; mark them unset.
engine->in_rgb_surf = VA_INVALID_ID;
engine->nv12_surf = VA_INVALID_ID;
engine->recon_surfs[0] = VA_INVALID_ID;
engine->recon_surfs[1] = VA_INVALID_ID;
engine->vpp_cfg = VA_INVALID_ID;
engine->vpp_ctx = VA_INVALID_ID;
engine->vpp_buf = VA_INVALID_ID;
engine->enc_cfg = VA_INVALID_ID;
engine->enc_ctx = VA_INVALID_ID;
engine->coded_buf = VA_INVALID_ID;
engine->src_rect.width = config->cap_width;
engine->src_rect.height = config->cap_height;
engine->dst_rect.width = config->width;
engine->dst_rect.height = config->height;
engine->drm_fd = open(config->device ? config->device : "/dev/dri/renderD128", O_RDWR);
if (engine->drm_fd < 0) {
fprintf(stderr, "[vaapi] failed to open %s\n", config->device ? config->device : "/dev/dri/renderD128");
free(engine);
return NULL;
}
engine->va_dpy = RT.vaGetDisplayDRM(engine->drm_fd);
if (!engine->va_dpy) {
close(engine->drm_fd);
free(engine);
return NULL;
}
int maj, min;
VAStatus st = RT.vaInitialize(engine->va_dpy, &maj, &min);
if (st != VA_STATUS_SUCCESS) {
fprintf(stderr, "[vaapi] vaInitialize failed: %s\n", RT.vaErrorStr(st));
close(engine->drm_fd);
free(engine);
return NULL;
}
// 1. NV12 intermediate + two Recon surfaces at the encode size (ping-pong: a P frame must
// not reference the surface its own reconstruction is being written into)
st = RT.vaCreateSurfaces(engine->va_dpy, VA_RT_FORMAT_YUV420, engine->width, engine->height, &engine->nv12_surf, 1, NULL, 0);
if (st != VA_STATUS_SUCCESS) {
vaapi_destroy(engine);
return NULL;
}
st = RT.vaCreateSurfaces(engine->va_dpy, VA_RT_FORMAT_YUV420, engine->width, engine->height, engine->recon_surfs, 2, NULL, 0);
if (st != VA_STATUS_SUCCESS) {
vaapi_destroy(engine);
return NULL;
}
// 2. Setup VPP (Video Processing Pipeline for Hardware RGB->NV12 with scaling)
uint32_t vpp_w = engine->cap_width > engine->width ? engine->cap_width : engine->width;
uint32_t vpp_h = engine->cap_height > engine->height ? engine->cap_height : engine->height;
st = RT.vaCreateConfig(engine->va_dpy, VAProfileNone, VAEntrypointVideoProc, NULL, 0, &engine->vpp_cfg);
if (st != VA_STATUS_SUCCESS) {
vaapi_destroy(engine);
return NULL;
}
st = RT.vaCreateContext(engine->va_dpy, engine->vpp_cfg, vpp_w, vpp_h, 0, NULL, 0, &engine->vpp_ctx);
if (st != VA_STATUS_SUCCESS) {
vaapi_destroy(engine);
return NULL;
}
// 3. Setup H.264 / HEVC EncSlice
VAProfile profile = engine->is_hevc ? VAProfileHEVCMain : VAProfileH264Main;
VAConfigAttrib enc_attrib[3];
enc_attrib[0].type = VAConfigAttribRTFormat;
enc_attrib[0].value = VA_RT_FORMAT_YUV420;
enc_attrib[1].type = VAConfigAttribRateControl;
// The engine writes SPS, PPS and slice headers itself (see BitWriter)
enc_attrib[2].type = VAConfigAttribEncPackedHeaders;
enc_attrib[2].value = VA_ENC_PACKED_HEADER_SEQUENCE | VA_ENC_PACKED_HEADER_SLICE;
// VBR like the FFmpeg VAAPI pipeline (target + maxrate), then CBR, then CQP
enc_attrib[1].value = VA_RC_VBR;
engine->rc_mode = VA_RC_VBR;
st = RT.vaCreateConfig(engine->va_dpy, profile, VAEntrypointEncSlice, enc_attrib, 3, &engine->enc_cfg);
if (st != VA_STATUS_SUCCESS) {
enc_attrib[1].value = VA_RC_CBR;
engine->rc_mode = VA_RC_CBR;
st = RT.vaCreateConfig(engine->va_dpy, profile, VAEntrypointEncSlice, enc_attrib, 3, &engine->enc_cfg);
}
if (st != VA_STATUS_SUCCESS) {
// Fallback to CQP if CBR unsupported
enc_attrib[1].value = VA_RC_CQP;
st = RT.vaCreateConfig(engine->va_dpy, profile, VAEntrypointEncSlice, enc_attrib, 3, &engine->enc_cfg);
if (st != VA_STATUS_SUCCESS) {
fprintf(stderr, "[vaapi] vaCreateConfig EncSlice failed: %s\n", RT.vaErrorStr(st));
vaapi_destroy(engine);
return NULL;
}
engine->rc_mode = VA_RC_CQP;
}
st = RT.vaCreateContext(engine->va_dpy, engine->enc_cfg, engine->width, engine->height, VA_PROGRESSIVE, engine->recon_surfs, 2, &engine->enc_ctx);
if (st != VA_STATUS_SUCCESS) {
fprintf(stderr, "[vaapi] vaCreateContext EncSlice failed: %s\n", RT.vaErrorStr(st));
vaapi_destroy(engine);
return NULL;
}
// Coded bitstream buffer: generous, a 20M-preset intra frame can burst past 4MB
st = RT.vaCreateBuffer(engine->va_dpy, engine->enc_ctx, VAEncCodedBufferType, 16 * 1024 * 1024, 1, NULL, &engine->coded_buf);
if (st != VA_STATUS_SUCCESS) {
vaapi_destroy(engine);
return NULL;
}
return engine;
}
int vaapi_import_dma_buffer(VaapiEngine* engine, int dma_fd, uint32_t stride) {
if (!engine || dma_fd < 0) return -1;
if (engine->in_rgb_surf != VA_INVALID_ID) {
RT.vaDestroySurfaces(engine->va_dpy, &engine->in_rgb_surf, 1);
engine->in_rgb_surf = VA_INVALID_ID;
}
if (engine->vpp_buf != VA_INVALID_ID) {
RT.vaDestroyBuffer(engine->va_dpy, engine->vpp_buf);
engine->vpp_buf = VA_INVALID_ID;
}
engine->dma_fd = dma_fd;
engine->stride = stride;
VADRMPRIMESurfaceDescriptor prime_desc;
memset(&prime_desc, 0, sizeof(prime_desc));
prime_desc.fourcc = VA_FOURCC_BGRX;
prime_desc.width = engine->cap_width;
prime_desc.height = engine->cap_height;
prime_desc.num_objects = 1;
prime_desc.objects[0].fd = dma_fd;
prime_desc.objects[0].size = stride * engine->cap_height;
prime_desc.objects[0].drm_format_modifier = 0;
prime_desc.num_layers = 1;
prime_desc.layers[0].drm_format = 0x34325258; // DRM_FORMAT_XRGB8888
prime_desc.layers[0].num_planes = 1;
prime_desc.layers[0].pitch[0] = stride;
VASurfaceAttrib attribs[2];
attribs[0].type = VASurfaceAttribMemoryType;
attribs[0].flags = VA_SURFACE_ATTRIB_SETTABLE;
attribs[0].value.type = VAGenericValueTypeInteger;
attribs[0].value.value.i = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME_2;
attribs[1].type = VASurfaceAttribExternalBufferDescriptor;
attribs[1].flags = VA_SURFACE_ATTRIB_SETTABLE;
attribs[1].value.type = VAGenericValueTypePointer;
attribs[1].value.value.p = &prime_desc;
VAStatus st = RT.vaCreateSurfaces(engine->va_dpy, VA_RT_FORMAT_RGB32, engine->cap_width, engine->cap_height, &engine->in_rgb_surf, 1, attribs, 2);
if (st != VA_STATUS_SUCCESS) {
fprintf(stderr, "[vaapi] vaCreateSurfaces DRM_PRIME_2 failed: %s\n", RT.vaErrorStr(st));
return -1;
}
if (vaapi_make_vpp_buffer(engine) != 0) {
fprintf(stderr, "[vaapi] vaCreateBuffer VPP failed\n");
return -2;
}
return 0;
}
int vaapi_upload_host_buffer(VaapiEngine* engine, const uint8_t* host_ptr, uint32_t pitch) {
if (!engine || !host_ptr) return -1;
if (engine->in_rgb_surf == VA_INVALID_ID) {
VAStatus st = RT.vaCreateSurfaces(engine->va_dpy, VA_RT_FORMAT_RGB32, engine->cap_width, engine->cap_height, &engine->in_rgb_surf, 1, NULL, 0);
if (st != VA_STATUS_SUCCESS) return -2;
if (vaapi_make_vpp_buffer(engine) != 0) return -3;
}
if (!engine->upload_ready) {
VAImageFormat fmt;
memset(&fmt, 0, sizeof(fmt));
fmt.fourcc = VA_FOURCC_BGRX;
fmt.byte_order = VA_LSB_FIRST;
fmt.bits_per_pixel = 32;
fmt.depth = 24;
fmt.red_mask = 0x00ff0000;
fmt.green_mask = 0x0000ff00;
fmt.blue_mask = 0x000000ff;
// vaDeriveImage on the RGB32 surface is refused by some drivers
VAStatus st = RT.vaCreateImage(engine->va_dpy, &fmt, engine->cap_width, engine->cap_height, &engine->upload_img);
if (st != VA_STATUS_SUCCESS) return -4;
engine->upload_ready = 1;
}
void* p = NULL;
VAStatus st = RT.vaMapBuffer(engine->va_dpy, engine->upload_img.buf, &p);
if (st != VA_STATUS_SUCCESS || !p) return -5;
uint32_t copy_bytes = engine->cap_width * 4;
for (uint32_t y = 0; y < engine->cap_height; y++) {
memcpy((uint8_t*)p + y * engine->upload_img.pitches[0], host_ptr + y * pitch, copy_bytes);
}
RT.vaUnmapBuffer(engine->va_dpy, engine->upload_img.buf);
st = RT.vaPutImage(engine->va_dpy, engine->in_rgb_surf, engine->upload_img.image_id,
0, 0, engine->cap_width, engine->cap_height, 0, 0, engine->cap_width, engine->cap_height);
if (st != VA_STATUS_SUCCESS) return -6;
return 0;
}
// Sequence-level misc parameter. AMD's Mesa driver feeds the VCN ring message from these;
// a CBR config that never gets its rate-control/HRD parameters submits garbage and the
// encode job wedges the ring until a GPU reset.
// Packed H.264 / HEVC headers. Current radeonsi (Mesa 26) writes neither SPS/PPS nor the NAL/slice header from the
// parameter buffers alone (output began 00 00 00 01 00 ..., no SPS), so the host writes them like FFmpeg does.
// Every field here must match the sequence/picture/slice parameters submitted with it.
typedef struct { uint8_t buf[128]; uint32_t bits; } BitWriter;
static void bw_put(BitWriter* w, uint32_t value, int n) {
for (int i = n - 1; i >= 0; i--, w->bits++) {
if (w->bits / 8 >= sizeof(w->buf)) return;
if ((value >> i) & 1) w->buf[w->bits / 8] |= 0x80 >> (w->bits % 8);
}
}
static void bw_ue(BitWriter* w, uint32_t value) {
int len = 0;
for (uint32_t v = value + 1; v > 1; v >>= 1) len++;
bw_put(w, 0, len);
bw_put(w, value + 1, len + 1);
}
static void bw_se(BitWriter* w, int32_t value) { bw_ue(w, value > 0 ? 2 * (uint32_t)value - 1 : 2 * (uint32_t)(-value)); }
static void bw_trailing(BitWriter* w) { bw_put(w, 1, 1); if (w->bits % 8) bw_put(w, 0, 8 - w->bits % 8); }
// Start code + NAL header; the payload follows unescaped, emulation prevention is applied by bw_escape.
static void bw_nal(BitWriter* w, int ref_idc, int type) { bw_put(w, 1, 32); bw_put(w, (ref_idc << 5) | type, 8); }
// Insert emulation prevention bytes after the start code + NAL header (byte 5 on). Returns the byte length.
static uint32_t bw_escape(const BitWriter* w, uint8_t* out, uint32_t cap) {
uint32_t len = (w->bits + 7) / 8, n = 0, zeros = 0;
for (uint32_t i = 0; i < len && n < cap; i++) {
if (i >= 5 && zeros >= 2 && w->buf[i] <= 3) { out[n++] = 3; zeros = 0; if (n >= cap) break; }
zeros = w->buf[i] == 0 && i >= 5 ? zeros + 1 : 0;
out[n++] = w->buf[i];
}
return n;
}
// Each NAL is escaped on its own, so the next NAL's start code stays intact.
static int vaapi_add_packed(VaapiEngine* engine, VAEncPackedHeaderType type, const BitWriter* nals, int count,
VABufferID* ids, int* n) {
uint8_t data[320];
uint32_t bytes = 0;
for (int i = 0; i < count; i++) bytes += bw_escape(&nals[i], data + bytes, sizeof(data) - bytes);
VAEncPackedHeaderParameterBuffer param = { .type = type, .bit_length = bytes * 8, .has_emulation_bytes = 1 };
if (RT.vaCreateBuffer(engine->va_dpy, engine->enc_ctx, VAEncPackedHeaderParameterBufferType, sizeof(param), 1, &param, &ids[*n]) != VA_STATUS_SUCCESS)
return -1;
(*n)++;
if (RT.vaCreateBuffer(engine->va_dpy, engine->enc_ctx, VAEncPackedHeaderDataBufferType, bytes, 1, data, &ids[*n]) != VA_STATUS_SUCCESS)
return -1;
(*n)++;
return 0;
}
static VABufferID vaapi_create_misc_buffer(VaapiEngine* engine, VAEncMiscParameterType type, const void* param, uint32_t param_size) {
uint8_t buf[sizeof(VAEncMiscParameterBuffer) + 64];
VAEncMiscParameterBuffer* misc = (VAEncMiscParameterBuffer*)buf;
misc->type = type;
memcpy(misc->data, param, param_size);
VABufferID id = VA_INVALID_ID;
if (RT.vaCreateBuffer(engine->va_dpy, engine->enc_ctx, VAEncMiscParameterBufferType,
sizeof(VAEncMiscParameterBuffer) + param_size, 1, buf, &id) != VA_STATUS_SUCCESS) {
return VA_INVALID_ID;
}
return id;
}
// Buffers for one encoded picture; every created ID is destroyed after the picture, success or not.
typedef struct {
VABufferID seq, misc[3], pic, slice, packed[6];
int misc_n, packed_n;
} EncBuffers;
static int vaapi_h264_params(VaapiEngine* engine, int is_idr, VASurfaceID curr_recon, VASurfaceID prev_recon, EncBuffers* b) {
uint32_t mb_w = (engine->width + 15) / 16;
uint32_t mb_h = (engine->height + 15) / 16;
uint32_t frame_num = is_idr ? 0 : engine->frame_num;
uint32_t crop_right = (mb_w * 16 - engine->width) / 2, crop_bottom = (mb_h * 16 - engine->height) / 2;
if (is_idr) {
VAEncSequenceParameterBufferH264 seq_p;
memset(&seq_p, 0, sizeof(seq_p));
seq_p.seq_parameter_set_id = 0;
seq_p.level_idc = 51;
seq_p.intra_period = engine->gop;
seq_p.intra_idr_period = engine->gop;
seq_p.ip_period = 1;
seq_p.bits_per_second = (uint32_t)engine->max_bps;
seq_p.max_num_ref_frames = 1;
seq_p.picture_width_in_mbs = mb_w;
seq_p.picture_height_in_mbs = mb_h;
seq_p.seq_fields.bits.direct_8x8_inference_flag = 1;
seq_p.seq_fields.bits.frame_mbs_only_flag = 1;
seq_p.seq_fields.bits.pic_order_cnt_type = 0;
seq_p.frame_cropping_flag = crop_right || crop_bottom;
seq_p.frame_crop_right_offset = crop_right;
seq_p.frame_crop_bottom_offset = crop_bottom;
// An IDR without its SPS is undecodable: fail rather than submit pic + slice alone
if (RT.vaCreateBuffer(engine->va_dpy, engine->enc_ctx, VAEncSequenceParameterBufferType, sizeof(seq_p), 1, &seq_p, &b->seq) != VA_STATUS_SUCCESS)
return -1;
// SPS: Main, level 5.1, frame_num and POC lsb 4 bits (both wrap & 15 above), one reference
BitWriter sps = {0};
bw_nal(&sps, 3, 7);
bw_put(&sps, 77, 8); bw_put(&sps, 0x40, 8); bw_put(&sps, 51, 8); // constraint_set1: Main
bw_ue(&sps, 0); bw_ue(&sps, 0); bw_ue(&sps, 0); bw_ue(&sps, 0); // sps id, log2_max_frame_num-4, poc type 0, log2_max_poc_lsb-4
bw_ue(&sps, 1); bw_put(&sps, 0, 1); // max_num_ref_frames, gaps
bw_ue(&sps, mb_w - 1); bw_ue(&sps, mb_h - 1);
bw_put(&sps, 1, 1); bw_put(&sps, 1, 1); // frame_mbs_only, direct_8x8_inference
bw_put(&sps, seq_p.frame_cropping_flag, 1);
if (seq_p.frame_cropping_flag) { bw_ue(&sps, 0); bw_ue(&sps, crop_right); bw_ue(&sps, 0); bw_ue(&sps, crop_bottom); }
bw_put(&sps, 0, 1); // no VUI
bw_trailing(&sps);
// PPS: CABAC, one L0 reference, QP 26, default deblocking
BitWriter pps = {0};
bw_nal(&pps, 3, 8);
bw_ue(&pps, 0); bw_ue(&pps, 0); bw_put(&pps, 1, 1); bw_put(&pps, 0, 1);
bw_ue(&pps, 0); bw_ue(&pps, 0); bw_ue(&pps, 0); // slice groups, num_ref_idx l0/l1 - 1
bw_put(&pps, 0, 1); bw_put(&pps, 0, 2); // weighted pred / bipred
bw_se(&pps, 0); bw_se(&pps, 0); bw_se(&pps, 0); // pic_init_qp-26, qs, chroma offset
bw_put(&pps, 0, 1); bw_put(&pps, 0, 1); bw_put(&pps, 0, 1); // deblocking control, constrained intra, redundant
bw_trailing(&pps);
// SPS and PPS go out as one Sequence header, like FFmpeg
BitWriter headers[2] = { sps, pps };
if (vaapi_add_packed(engine, VAEncPackedHeaderSequence, headers, 2, b->packed, &b->packed_n) != 0) return -1;
}
VAEncPictureParameterBufferH264 pic_p;
memset(&pic_p, 0, sizeof(pic_p));
pic_p.CurrPic.picture_id = curr_recon;
pic_p.CurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
pic_p.CurrPic.frame_idx = frame_num;
pic_p.CurrPic.TopFieldOrderCnt = pic_p.CurrPic.BottomFieldOrderCnt = (2 * frame_num) & 15;
for (int i = 0; i < 16; i++) pic_p.ReferenceFrames[i].picture_id = VA_INVALID_ID;
if (!is_idr) {
pic_p.ReferenceFrames[0].picture_id = prev_recon;
pic_p.ReferenceFrames[0].flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
pic_p.ReferenceFrames[0].frame_idx = (frame_num - 1) & 15;
pic_p.ReferenceFrames[0].TopFieldOrderCnt = pic_p.ReferenceFrames[0].BottomFieldOrderCnt = (2 * (frame_num - 1)) & 15;
}
pic_p.coded_buf = engine->coded_buf;
pic_p.frame_num = frame_num;
pic_p.pic_init_qp = 26;
pic_p.seq_parameter_set_id = 0;
pic_p.pic_parameter_set_id = 0;
pic_p.pic_fields.bits.idr_pic_flag = is_idr ? 1 : 0;
pic_p.pic_fields.bits.reference_pic_flag = 1;
pic_p.pic_fields.bits.entropy_coding_mode_flag = 1;
pic_p.pic_fields.bits.deblocking_filter_control_present_flag = 0;
if (RT.vaCreateBuffer(engine->va_dpy, engine->enc_ctx, VAEncPictureParameterBufferType, sizeof(pic_p), 1, &pic_p, &b->pic) != VA_STATUS_SUCCESS)
return -1;
VAEncSliceParameterBufferH264 slice_p;
memset(&slice_p, 0, sizeof(slice_p));
slice_p.macroblock_address = 0;
slice_p.num_macroblocks = mb_w * mb_h;
slice_p.macroblock_info = VA_INVALID_ID;
slice_p.slice_type = is_idr ? 2 : 0; // 2 = I-Slice, 0 = P-Slice
slice_p.direct_spatial_mv_pred_flag = 1;
// Consecutive IDRs must differ in idr_pic_id (7.4.3); flip it on every IDR
if (is_idr) engine->idr_pic_id ^= 1;
slice_p.idr_pic_id = engine->idr_pic_id;
slice_p.pic_order_cnt_lsb = pic_p.CurrPic.TopFieldOrderCnt;
for (int i = 0; i < 32; i++) slice_p.RefPicList0[i].picture_id = VA_INVALID_ID;
if (!is_idr) {
slice_p.RefPicList0[0].picture_id = prev_recon;
slice_p.RefPicList0[0].flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
slice_p.RefPicList0[0].frame_idx = (frame_num - 1) & 15;
}
if (RT.vaCreateBuffer(engine->va_dpy, engine->enc_ctx, VAEncSliceParameterBufferType, sizeof(slice_p), 1, &slice_p, &b->slice) != VA_STATUS_SUCCESS)
return -1;
// Slice header including cabac_alignment_one_bit, byte aligned and escaped as FFmpeg submits it
BitWriter sh = {0};
bw_nal(&sh, 3, is_idr ? 5 : 1);
bw_ue(&sh, 0); bw_ue(&sh, is_idr ? 7 : 5); bw_ue(&sh, 0); // first_mb, slice_type I/P (all slices), pps id
bw_put(&sh, frame_num & 15, 4);
if (is_idr) bw_ue(&sh, slice_p.idr_pic_id);
bw_put(&sh, slice_p.pic_order_cnt_lsb & 15, 4);
if (!is_idr) { bw_put(&sh, 0, 1); bw_put(&sh, 0, 1); } // num_ref_idx override, ref list modification
if (is_idr) { bw_put(&sh, 0, 1); bw_put(&sh, 0, 1); } // no_output_of_prior_pics, long_term_reference
else bw_put(&sh, 0, 1); // adaptive_ref_pic_marking
if (!is_idr) bw_ue(&sh, 0); // cabac_init_idc
bw_se(&sh, 0); // slice_qp_delta (no deblocking fields)
while (sh.bits % 8) bw_put(&sh, 1, 1);
return vaapi_add_packed(engine, VAEncPackedHeaderSlice, &sh, 1, b->packed, &b->packed_n);
}
// HEVC NAL header: forbidden 0, nal_unit_type, nuh_layer_id 0, nuh_temporal_id_plus1 1.
static void bw_hevc_nal(BitWriter* w, int type) { bw_put(w, 1, 32); bw_put(w, (type << 9) | 1, 16); }
// profile_tier_level(1, 0): Main, Main 10 compatible like FFmpeg, progressive frames only.
static void bw_hevc_ptl(BitWriter* w, uint32_t level) {
bw_put(w, 0, 2); bw_put(w, 0, 1); bw_put(w, 1, 5); // profile space, tier, general_profile_idc Main
bw_put(w, 0x60000000, 32); // compatibility flags 1 and 2
bw_put(w, 1, 1); bw_put(w, 0, 1); bw_put(w, 1, 1); bw_put(w, 1, 1); // progressive, interlaced, non_packed, frame_only
bw_put(w, 0, 32); bw_put(w, 0, 12); // reserved_zero_43bits + inbld_flag
bw_put(w, level, 8);
}
// HEVC Main with FFmpeg's hevc_vaapi parameters on radeonsi (LIBVA_TRACE + trace_headers of its output):
// 64x64 CTB, 8x8 min CB, 4x4..32x32 TB, AMP and SAO on, one short-term reference, POC lsb 12 bits.
// FFmpeg also sends VA max_transform_hierarchy_depth 0 / diff_cu_qp_delta_depth 3 while its SPS/PPS
// say 4 / 0; both copies are kept as FFmpeg has them, since that is what decodes on this driver.
static int vaapi_hevc_params(VaapiEngine* engine, int is_idr, VASurfaceID curr_recon, VASurfaceID prev_recon, EncBuffers* b) {
// pic size must be a multiple of the 8x8 min CB; the conformance window (in chroma units) crops the rest
uint32_t w = (engine->width + 7) & ~7u, h = (engine->height + 7) & ~7u;
uint32_t crop_right = (w - engine->width) / 2, crop_bottom = (h - engine->height) / 2;
uint32_t poc = is_idr ? 0 : engine->since_idr;
// Level from luma picture size and sample rate (Table A.8): 4.1, 5.0, 5.1, 5.2
uint64_t size = (uint64_t)w * h, rate = size * engine->fps;
uint32_t level = rate > 534773760 ? 156 : rate > 267386880 ? 153 : (size > 2228224 || rate > 133693440) ? 150 : 123;
if (is_idr) {
VAEncSequenceParameterBufferHEVC seq_p;
memset(&seq_p, 0, sizeof(seq_p));
seq_p.general_profile_idc = 1;
seq_p.general_level_idc = level;
seq_p.intra_period = engine->gop;
seq_p.intra_idr_period = engine->gop;
seq_p.ip_period = 1;
seq_p.bits_per_second = (uint32_t)engine->max_bps;
seq_p.pic_width_in_luma_samples = w;
seq_p.pic_height_in_luma_samples = h;
seq_p.seq_fields.bits.chroma_format_idc = 1;
seq_p.seq_fields.bits.amp_enabled_flag = 1;
seq_p.seq_fields.bits.sample_adaptive_offset_enabled_flag = 1;
seq_p.log2_diff_max_min_luma_coding_block_size = 3;
seq_p.log2_diff_max_min_transform_block_size = 3;
if (RT.vaCreateBuffer(engine->va_dpy, engine->enc_ctx, VAEncSequenceParameterBufferType, sizeof(seq_p), 1, &seq_p, &b->seq) != VA_STATUS_SUCCESS)
return -1;
BitWriter vps = {0};
bw_hevc_nal(&vps, 32);
bw_put(&vps, 0, 4); bw_put(&vps, 1, 1); bw_put(&vps, 1, 1); // vps id, base layer internal / available
bw_put(&vps, 0, 6); bw_put(&vps, 0, 3); bw_put(&vps, 1, 1); // max layers-1, max sub layers-1, temporal id nesting
bw_put(&vps, 0xffff, 16);
bw_hevc_ptl(&vps, level);
bw_put(&vps, 0, 1); bw_ue(&vps, 1); bw_ue(&vps, 0); bw_ue(&vps, 0); // ordering info: dec buffering-1, reorder, latency
bw_put(&vps, 0, 6); bw_ue(&vps, 0); // max layer id, num layer sets-1
bw_put(&vps, 0, 1); bw_put(&vps, 0, 1); // no timing info, no extension
bw_trailing(&vps);
BitWriter sps = {0};
bw_hevc_nal(&sps, 33);
bw_put(&sps, 0, 4); bw_put(&sps, 0, 3); bw_put(&sps, 1, 1); // vps id, max sub layers-1, temporal id nesting
bw_hevc_ptl(&sps, level);
bw_ue(&sps, 0); bw_ue(&sps, 1); bw_ue(&sps, w); bw_ue(&sps, h); // sps id, 4:2:0, size
bw_put(&sps, crop_right || crop_bottom, 1);
if (crop_right || crop_bottom) { bw_ue(&sps, 0); bw_ue(&sps, crop_right); bw_ue(&sps, 0); bw_ue(&sps, crop_bottom); }
bw_ue(&sps, 0); bw_ue(&sps, 0); bw_ue(&sps, 8); // 8-bit luma/chroma, log2_max_poc_lsb-4
bw_put(&sps, 0, 1); bw_ue(&sps, 1); bw_ue(&sps, 0); bw_ue(&sps, 0);
bw_ue(&sps, 0); bw_ue(&sps, 3); bw_ue(&sps, 0); bw_ue(&sps, 3); // CB 8..64, TB 4..32
bw_ue(&sps, 4); bw_ue(&sps, 4); // max transform hierarchy depth inter/intra
bw_put(&sps, 0, 1); bw_put(&sps, 1, 1); bw_put(&sps, 1, 1); bw_put(&sps, 0, 1); // scaling list, AMP, SAO, PCM
bw_ue(&sps, 0); bw_put(&sps, 0, 1); // RPS in slice headers, no long-term refs
bw_put(&sps, 0, 1); bw_put(&sps, 0, 1); // temporal MVP, strong intra smoothing
bw_put(&sps, 0, 1); bw_put(&sps, 0, 1); // no VUI, no extension
bw_trailing(&sps);
BitWriter pps = {0};
bw_hevc_nal(&pps, 34);
bw_ue(&pps, 0); bw_ue(&pps, 0);
bw_put(&pps, 0, 1); bw_put(&pps, 0, 1); bw_put(&pps, 0, 3); // dependent slices, output flag, extra bits
bw_put(&pps, 0, 1); bw_put(&pps, 0, 1); // sign data hiding, cabac init present
bw_ue(&pps, 0); bw_ue(&pps, 0); bw_se(&pps, 0); // num_ref_idx l0/l1 default-1, init_qp-26
bw_put(&pps, 0, 1); bw_put(&pps, 1, 1); bw_put(&pps, 1, 1); // constrained intra, transform skip, cu_qp_delta
bw_ue(&pps, 0); bw_se(&pps, 0); bw_se(&pps, 0); // diff_cu_qp_delta_depth, cb/cr offsets
bw_put(&pps, 0, 1); bw_put(&pps, 0, 1); bw_put(&pps, 0, 1); // slice chroma offsets, weighted pred/bipred
bw_put(&pps, 0, 1); bw_put(&pps, 0, 1); bw_put(&pps, 0, 1); // transquant bypass, tiles, entropy sync
bw_put(&pps, 1, 1); bw_put(&pps, 0, 1); bw_put(&pps, 0, 1); // loop filter across slices, deblocking ctl, scaling
bw_put(&pps, 0, 1); bw_ue(&pps, 0); bw_put(&pps, 0, 1); bw_put(&pps, 0, 1); // list mod, merge level-2, ext
bw_trailing(&pps);
BitWriter headers[3] = { vps, sps, pps };
if (vaapi_add_packed(engine, VAEncPackedHeaderSequence, headers, 3, b->packed, &b->packed_n) != 0) return -1;
}
VAEncPictureParameterBufferHEVC pic_p;
memset(&pic_p, 0, sizeof(pic_p));
pic_p.decoded_curr_pic.picture_id = curr_recon;
pic_p.decoded_curr_pic.pic_order_cnt = poc;
for (int i = 0; i < 15; i++) {
pic_p.reference_frames[i].picture_id = VA_INVALID_ID;
pic_p.reference_frames[i].flags = VA_PICTURE_HEVC_INVALID;
}
if (!is_idr) {
pic_p.reference_frames[0].picture_id = prev_recon;
pic_p.reference_frames[0].pic_order_cnt = poc - 1;
pic_p.reference_frames[0].flags = VA_PICTURE_HEVC_RPS_ST_CURR_BEFORE;
}
pic_p.coded_buf = engine->coded_buf;
pic_p.collocated_ref_pic_index = 0xff;
pic_p.pic_init_qp = 30;
pic_p.diff_cu_qp_delta_depth = 3;
pic_p.nal_unit_type = is_idr ? 19 : 1; // IDR_W_RADL / TRAIL_R
pic_p.pic_fields.bits.idr_pic_flag = is_idr;
pic_p.pic_fields.bits.coding_type = is_idr ? 1 : 2;
pic_p.pic_fields.bits.reference_pic_flag = 1;
pic_p.pic_fields.bits.transform_skip_enabled_flag = 1;
pic_p.pic_fields.bits.cu_qp_delta_enabled_flag = 1;
pic_p.pic_fields.bits.pps_loop_filter_across_slices_enabled_flag = 1;
if (RT.vaCreateBuffer(engine->va_dpy, engine->enc_ctx, VAEncPictureParameterBufferType, sizeof(pic_p), 1, &pic_p, &b->pic) != VA_STATUS_SUCCESS)
return -1;
VAEncSliceParameterBufferHEVC slice_p;
memset(&slice_p, 0, sizeof(slice_p));
slice_p.num_ctu_in_slice = ((w + 63) / 64) * ((h + 63) / 64);
slice_p.slice_type = is_idr ? 2 : 1; // I / P
for (int i = 0; i < 15; i++) {
slice_p.ref_pic_list0[i].picture_id = slice_p.ref_pic_list1[i].picture_id = VA_INVALID_ID;
slice_p.ref_pic_list0[i].flags = slice_p.ref_pic_list1[i].flags = VA_PICTURE_HEVC_INVALID;
}
if (!is_idr) slice_p.ref_pic_list0[0] = pic_p.reference_frames[0];
slice_p.max_num_merge_cand = 5;
slice_p.slice_fields.bits.last_slice_of_pic_flag = 1;
slice_p.slice_fields.bits.slice_sao_luma_flag = 1;
slice_p.slice_fields.bits.slice_sao_chroma_flag = 1;
slice_p.slice_fields.bits.collocated_from_l0_flag = !is_idr;
if (RT.vaCreateBuffer(engine->va_dpy, engine->enc_ctx, VAEncSliceParameterBufferType, sizeof(slice_p), 1, &slice_p, &b->slice) != VA_STATUS_SUCCESS)
return -1;
BitWriter sh = {0};
bw_hevc_nal(&sh, pic_p.nal_unit_type);
bw_put(&sh, 1, 1); // first slice segment
if (is_idr) bw_put(&sh, 0, 1); // no_output_of_prior_pics
bw_ue(&sh, 0); bw_ue(&sh, slice_p.slice_type); // pps id, slice type
if (!is_idr) {
bw_put(&sh, poc & 0xfff, 12);
bw_put(&sh, 0, 1); // RPS coded here: one previous picture
bw_ue(&sh, 1); bw_ue(&sh, 0); bw_ue(&sh, 0); bw_put(&sh, 1, 1); // negative 1, positive 0, delta-1 0, used
}
bw_put(&sh, 1, 1); bw_put(&sh, 1, 1); // SAO luma, chroma
if (!is_idr) { bw_put(&sh, 0, 1); bw_ue(&sh, 0); } // num_ref_idx override, five_minus_max_merge
bw_se(&sh, 0); // slice_qp_delta (the driver writes the real one)
bw_put(&sh, 0, 1); // loop filter across slices
bw_put(&sh, 1, 1); // byte_alignment()
while (sh.bits % 8) bw_put(&sh, 0, 1);
return vaapi_add_packed(engine, VAEncPackedHeaderSlice, &sh, 1, b->packed, &b->packed_n);
}
int vaapi_encode_frame(
VaapiEngine* engine,
int force_idr,
const uint8_t** out_data,
uint32_t* out_len
) {
if (out_data) *out_data = NULL;
if (out_len) *out_len = 0;
if (!engine || engine->in_rgb_surf == VA_INVALID_ID) return -1;
int is_idr = force_idr || engine->idr_pending || engine->frame_count == 0 || engine->since_idr >= engine->gop;
// Recon ping-pong: a P frame must reference the previous recon surface, not its own
VASurfaceID curr_recon = engine->recon_surfs[engine->frame_count & 1];
VASurfaceID prev_recon = engine->recon_surfs[(engine->frame_count + 1) & 1];
// 1. Hardware VPP Color Conversion + scaling (RGB -> NV12 entirely inside VRAM)
VAStatus st = RT.vaBeginPicture(engine->va_dpy, engine->vpp_ctx, engine->nv12_surf);
if (st != VA_STATUS_SUCCESS) return -10;
st = RT.vaRenderPicture(engine->va_dpy, engine->vpp_ctx, &engine->vpp_buf, 1);
if (st != VA_STATUS_SUCCESS) return -10;
st = RT.vaEndPicture(engine->va_dpy, engine->vpp_ctx);
if (st != VA_STATUS_SUCCESS) return -10;
// 2. Rate control (every IDR) and the codec's sequence / picture / slice parameters + packed headers
EncBuffers b = { .seq = VA_INVALID_ID, .pic = VA_INVALID_ID, .slice = VA_INVALID_ID };
int ret = -11;
if (is_idr) {
VAEncMiscParameterRateControl rc;
memset(&rc, 0, sizeof(rc));
// VBR: bits_per_second is the ceiling (maxrate) and target_percentage the average (#33)
rc.bits_per_second = (uint32_t)engine->max_bps;
rc.target_percentage = engine->rc_mode == VA_RC_VBR ? (uint32_t)(engine->target_bps * 100 / engine->max_bps) : 100;
rc.window_size = 1000;
rc.initial_qp = 28;
rc.min_qp = 16;
rc.max_qp = 44;
rc.rc_flags.bits.disable_frame_skip = 1; // a remote desktop must not drop frames silently
b.misc[b.misc_n++] = vaapi_create_misc_buffer(engine, VAEncMiscParameterTypeRateControl, &rc, sizeof(rc));
VAEncMiscParameterHRD hrd;
memset(&hrd, 0, sizeof(hrd));
hrd.buffer_size = (uint32_t)engine->buffer_bits; // the viewer preset's bufsize
hrd.initial_buffer_fullness = hrd.buffer_size / 2;
b.misc[b.misc_n++] = vaapi_create_misc_buffer(engine, VAEncMiscParameterTypeHRD, &hrd, sizeof(hrd));
// numerator in the low 16 bits, denominator in the high 16 bits
VAEncMiscParameterFrameRate fr = { .framerate = (1u << 16) | engine->fps };
b.misc[b.misc_n++] = vaapi_create_misc_buffer(engine, VAEncMiscParameterTypeFrameRate, &fr, sizeof(fr));
}
if ((engine->is_hevc ? vaapi_hevc_params : vaapi_h264_params)(engine, is_idr, curr_recon, prev_recon, &b) != 0) goto done;
// 3. Render and Encode on AMD VCN hardware
VABufferID render_b[12];
int render_n = 0;
// FFmpeg's order: sequence, rate control misc, picture, packed headers (parameter sets, slice header), slice.
// With misc first, radeonsi (Mesa 26) ignored the VBR target: the 1M preset ran at 1.6 Mbps, 0.96 after.
if (b.seq != VA_INVALID_ID) render_b[render_n++] = b.seq;
for (int i = 0; i < b.misc_n; i++) {
if (b.misc[i] != VA_INVALID_ID) render_b[render_n++] = b.misc[i];
}
render_b[render_n++] = b.pic;
for (int i = 0; i < b.packed_n; i++) render_b[render_n++] = b.packed[i];
render_b[render_n++] = b.slice;
st = RT.vaBeginPicture(engine->va_dpy, engine->enc_ctx, engine->nv12_surf);
if (st != VA_STATUS_SUCCESS) goto done;
st = RT.vaRenderPicture(engine->va_dpy, engine->enc_ctx, render_b, render_n);
if (st != VA_STATUS_SUCCESS) goto done;
st = RT.vaEndPicture(engine->va_dpy, engine->enc_ctx);
if (st != VA_STATUS_SUCCESS) goto done;
st = RT.vaSyncSurface(engine->va_dpy, curr_recon);
if (st != VA_STATUS_SUCCESS) goto done;
// 4. Map Bitstream Output; an access unit may arrive as several linked segments
VACodedBufferSegment *segment = NULL;
st = RT.vaMapBuffer(engine->va_dpy, engine->coded_buf, (void**)&segment);
if (st != VA_STATUS_SUCCESS || !segment) goto done;
uint32_t total = 0;
for (VACodedBufferSegment *seg = segment; seg; seg = seg->next) total += seg->size;
if (total > 0 && !(segment->status & VA_CODED_BUF_STATUS_SLICE_OVERFLOW_MASK)) {
if (total > engine->output_staging_cap) {
engine->output_staging_cap = total + 64 * 1024;
engine->output_staging = (uint8_t*)realloc(engine->output_staging, engine->output_staging_cap);
}
uint32_t off = 0;
for (VACodedBufferSegment *seg = segment; seg; seg = seg->next) {
memcpy(engine->output_staging + off, seg->buf, seg->size);
off += seg->size;
}
*out_data = engine->output_staging;
*out_len = total;
}
RT.vaUnmapBuffer(engine->va_dpy, engine->coded_buf);
engine->frame_num = is_idr ? 1 : ((engine->frame_num + 1) & 15);
engine->since_idr = is_idr ? 1 : engine->since_idr + 1;
engine->frame_count++;
ret = 0;
done:
// No bitstream (failure, empty or slice overflow): the next frame_num / POC would skip, so restart with an IDR
engine->idr_pending = ret != 0 || *out_len == 0;
for (int i = 0; i < b.misc_n; i++) {
if (b.misc[i] != VA_INVALID_ID) RT.vaDestroyBuffer(engine->va_dpy, b.misc[i]);
}
if (b.seq != VA_INVALID_ID) RT.vaDestroyBuffer(engine->va_dpy, b.seq);
if (b.pic != VA_INVALID_ID) RT.vaDestroyBuffer(engine->va_dpy, b.pic);
if (b.slice != VA_INVALID_ID) RT.vaDestroyBuffer(engine->va_dpy, b.slice);
for (int i = 0; i < b.packed_n; i++) RT.vaDestroyBuffer(engine->va_dpy, b.packed[i]);
return ret;
}
void vaapi_destroy(VaapiEngine* engine) {
if (!engine) return;
if (engine->coded_buf != VA_INVALID_ID) RT.vaDestroyBuffer(engine->va_dpy, engine->coded_buf);
if (engine->enc_ctx != VA_INVALID_ID) RT.vaDestroyContext(engine->va_dpy, engine->enc_ctx);
if (engine->enc_cfg != VA_INVALID_ID) RT.vaDestroyConfig(engine->va_dpy, engine->enc_cfg);
if (engine->vpp_buf != VA_INVALID_ID) RT.vaDestroyBuffer(engine->va_dpy, engine->vpp_buf);
if (engine->vpp_ctx != VA_INVALID_ID) RT.vaDestroyContext(engine->va_dpy, engine->vpp_ctx);
if (engine->vpp_cfg != VA_INVALID_ID) RT.vaDestroyConfig(engine->va_dpy, engine->vpp_cfg);
if (engine->upload_ready) RT.vaDestroyImage(engine->va_dpy, engine->upload_img.image_id);
if (engine->in_rgb_surf != VA_INVALID_ID) RT.vaDestroySurfaces(engine->va_dpy, &engine->in_rgb_surf, 1);
if (engine->nv12_surf != VA_INVALID_ID) RT.vaDestroySurfaces(engine->va_dpy, &engine->nv12_surf, 1);
if (engine->recon_surfs[0] != VA_INVALID_ID || engine->recon_surfs[1] != VA_INVALID_ID) {
RT.vaDestroySurfaces(engine->va_dpy, engine->recon_surfs, 2);
}
if (engine->va_dpy) RT.vaTerminate(engine->va_dpy);
if (engine->drm_fd >= 0) close(engine->drm_fd);
if (engine->output_staging) free(engine->output_staging);
free(engine);
}
+49
View File
@@ -0,0 +1,49 @@
#ifndef UU_VAAPI_ENGINE_H
#define UU_VAAPI_ENGINE_H
#include <stdint.h>
#include <stddef.h>
#ifdef __cplusplus
extern "C" {
#endif
typedef struct VaapiEngine VaapiEngine;
typedef struct {
uint32_t width; // encode (output) size
uint32_t height;
uint32_t cap_width; // capture (input) size; VPP scales cap -> width
uint32_t cap_height;
uint32_t fps;
uint32_t gop;
int64_t target_bps;
int64_t max_bps; // VBR ceiling (viewer preset maxrate)
int64_t buffer_bits; // HRD buffer size
int is_hevc;
const char *device; // DRM render node (UU_VAAPI_DEVICE); NULL = /dev/dri/renderD128
} VaapiConfig;
VaapiEngine* vaapi_create(const VaapiConfig* config);
// Import a DRI3 / GBM DMA-BUF fd as the zero-copy input surface
int vaapi_import_dma_buffer(VaapiEngine* engine, int dma_fd, uint32_t stride);
// Host fallback: upload ARGB host memory if DMA-BUF is unavailable
int vaapi_upload_host_buffer(VaapiEngine* engine, const uint8_t* host_ptr, uint32_t pitch);
// Encode one frame
int vaapi_encode_frame(
VaapiEngine* engine,
int force_idr,
const uint8_t** out_data,
uint32_t* out_len
);
void vaapi_destroy(VaapiEngine* engine);
#ifdef __cplusplus
}
#endif
#endif // UU_VAAPI_ENGINE_H
+20
View File
@@ -0,0 +1,20 @@
官方控制端能选「真彩」(4:4:4) 了;自行编译的 `inproc-av` 新增进程内 NVENC 与 AMD/Intel VAAPI。/ The official viewer can select true color (4:4:4); self-built `inproc-av` gains in-process NVENC and VAAPI.
### 修复 / Fixed
- **官方控制端里「真彩」不可选或打开后马上被切回**(#77、#79):设备能力里 HEVC/4:4:4 等格式的宽高原来填的是 0x0,官方控制端会把这样的条目当作不支持。现在每种格式都带上最大尺寸 3840x2160。
已实机验收:官方 Mac 控制端连 X11 + NVIDIA 主机,打开真彩后切到 HEVC 4:4:4,持续 30 秒以上没有被切回,也没有请求关键帧(PLI)。
- The device capability listed HEVC/4:4:4 formats as 0x0, which the official viewer treats as unsupported; each format now
carries its maximum size. Verified with the official Mac viewer: true color stays on (HEVC 4:4:4) for 30+ s, no PLI.
### 新增(需自行编译,预编译包不受影响)/ Added (self-built only; release packages unchanged)
- `--features inproc-av`(仅 X11):
- **NVENC 进程内编码**(#75、#77,感谢 @ferstar):直接读取 XShm,支持 4:4:4,不再经过 FFmpeg 进程;失败时依次回退到 FFmpeg 的 NVENC 管线、CPU。我方 NVIDIA 机器未单独验收这条进程内路径。
- **VAAPI 进程内编码**(#80、#81,感谢 @ferstar):通过 DRI3 DMA-BUF 零拷贝取屏(没有 DRI3 时改用 XShm 上传),经 VPP 转成 NV12 后送入硬件编码,支持 H.264 和 H.265。贡献者已在 AMD HawkPoint、Cezanne 上用官方 Mac 控制端实机验收;Intel 与 4K 尚未验证。需设置 `UU_VIDEO_ENCODER=vaapi`。
- libcuda、libva 都在运行时加载;编译所需的头文件见 README。
- In-process NVENC (XShm, 4:4:4) and VAAPI (DRI3 DMA-BUF zero copy, H.264/H.265) under `--features inproc-av` on X11,
each falling back to the FFmpeg process pipeline. VAAPI accepted by the contributor on AMD HawkPoint/Cezanne; Intel and 4K untested.
### 下载 / Download
`uu-remote-linux_0.2.22_<amd64|arm64>.deb`,或 `uu-remote-linux-v0.2.22-<x86_64|aarch64>-linux.tar.gz`,或 `docker pull ghcr.io/vallierdev/uu-remote-linux:v0.2.22`(x86_64)。
**非官方项目,与网易无关。** / **Unofficial, not affiliated with NetEase.**
+1 -10
View File
@@ -1,7 +1,6 @@
//! Desktop sound for the viewer's 声音 menu: the default output's monitor (PulseAudio or
//! PipeWire-pulse) is encoded to Opus and sent on the negotiated audio m-line.
//! Encoding runs in this process with the `inproc-av` feature (unless `UU_FFMPEG_ROOT` is set),
//! otherwise an FFmpeg child process does it.
//! An FFmpeg child process does the encoding.
//! PipeWire monitors ignore the sink's mute by default, so 被控端静音运行 keeps the stream audible.
//! `UU_AUDIO=0` disables it; `UU_AUDIO_SOURCE` picks another Pulse source.
@@ -57,14 +56,6 @@ fn read_packets(stdout: impl Read, tx: &tokio::sync::mpsc::Sender<bytes::Bytes>)
fn start(source: &str, tx: tokio::sync::mpsc::Sender<bytes::Bytes>)
-> Result<(Stopper, tokio::task::JoinHandle<Result<(), String>>), String> {
let stop = Arc::new(AtomicBool::new(false));
#[cfg(feature = "inproc-av")]
if crate::inproc_encoder::selected() {
let job = tokio::task::spawn_blocking({
let (stop, source) = (stop.clone(), source.to_owned());
move || crate::inproc_encoder::run_audio(&source, tx, stop)
});
return Ok((Stopper { stop, child: None }, job));
}
let mut child = spawn_encoder(source)?;
let stdout = child.stdout.take().ok_or("Encoder stdout missing")?;
let job = tokio::task::spawn_blocking(move || read_packets(stdout, &tx));
+4 -3
View File
@@ -35,12 +35,13 @@ pub fn h264_device_capability(
/// The payload sits under `data.device_capability` with `type` beside it. The earlier flat shape
/// (`type` at the top, fields directly in `data`) was not recognised by the iPhone client, which
/// then kept quality and chroma locked at their defaults (#49). Only 8-bit and the chroma formats
/// the encoder really produces are listed; the first entry carries the maximum size as officially.
/// the encoder really produces are listed, each with its maximum size: the official viewer treats a
/// 0x0 entry as unsupported (#77; the 4.37 host we copied had a GTX 660 without HEVC/4:4:4).
pub fn device_capability(client_id: &str, ice_id: &str, display_id: i32, fps: i32, codec_impl: i32, hevc: bool, full_chroma: bool) -> Value {
let mut codecs = Vec::new();
for codec in [VIDEO_CODEC_H264].into_iter().chain(hevc.then_some(2)) {
for chroma in [1].into_iter().chain(full_chroma.then_some(3)) {
let (width, height) = if codecs.is_empty() { (3840, 2160) } else { (0, 0) };
let (width, height) = (3840, 2160);
codecs.push(json!({"video_codec": codec, "chroma_sampling": chroma, "codec_impl": codec_impl,
"bit_depth": 8, "width": width, "height": height}));
}
@@ -86,7 +87,7 @@ mod tests {
let caps = &payload["data"]["device_capability"]["video_codec_capability"];
assert_eq!(caps.as_array().unwrap().len(), 4);
assert_eq!(caps[0], json!({"video_codec":1,"chroma_sampling":1,"codec_impl":5,"bit_depth":8,"width":3840,"height":2160}));
assert_eq!(caps[3], json!({"video_codec":2,"chroma_sampling":3,"codec_impl":5,"bit_depth":8,"width":0,"height":0}));
assert_eq!(caps[3], json!({"video_codec":2,"chroma_sampling":3,"codec_impl":5,"bit_depth":8,"width":3840,"height":2160}));
assert_eq!(payload["data"]["device_capability"]["display_info"][0]["hdr"], -1);
let h264_420 = device_capability("c", "i", 0, 60, 5, false, false);
assert_eq!(h264_420["data"]["device_capability"]["video_codec_capability"].as_array().unwrap().len(), 1);
+10
View File
@@ -70,6 +70,16 @@ impl DamageCapture {
pub fn size(&self) -> (u32, u32) { (self.width as u32, self.height as u32) }
pub fn shm_raw_ptr(&self) -> *mut u8 { self.shm_ptr }
pub fn shm_size(&self) -> usize { self.len }
pub fn refresh(&mut self) -> Result<(), String> {
self.conn.shm_get_image(self.root, 0, 0, self.width, self.height, !0, Z_PIXMAP, self.seg, 0)
.map_err(text)?.reply().map_err(text)?;
Ok(())
}
/// Waits up to `timeout` for the screen to change. Returns `true` on change, `false` on timeout.
pub fn wait(&mut self, timeout: Duration) -> Result<bool, String> {
let deadline = Instant::now() + timeout;
+671 -397
View File
File diff suppressed because it is too large. Load diff
+33 -13
View File
@@ -198,11 +198,9 @@ impl Encoder {
fn nvidia_present() -> bool {
std::path::Path::new("/proc/driver/nvidia/version").exists() || std::path::Path::new("/dev/nvidia0").exists()
}
/// 真彩 (4:4:4) needs the FFmpeg process pipeline with NVENC or libx264 (h264_vaapi and the
/// in-process encoder are 4:2:0 only).
/// 真彩 (4:4:4): NVENC (in-process or FFmpeg) and libx264; h264_vaapi is 4:2:0 only.
/// A GPU without 4:4:4 NVENC falls back to libx264, which keeps 4:4:4.
pub fn full_chroma_available() -> bool {
#[cfg(feature = "inproc-av")]
if crate::inproc_encoder::selected() { return false; }
Encoder::requested() != Encoder::Vaapi
}
@@ -506,17 +504,33 @@ async fn read_encoder(
) -> Result<(), String> {
// The sequence is shared so a CPU fallback or settings restart continues it (skip statistics stay correct).
let encoder = Encoder::requested();
if encoder == Encoder::Cpu { return read_encoder_with(size,display,portal,latest,Encoder::Cpu,sequence,settings,idr,budget).await; }
match read_encoder_with(size,display.clone(),portal.clone(),latest.clone(),encoder,sequence.clone(),settings,idr.clone(),budget.clone()).await {
if encoder == Encoder::Cpu { return read_encoder_with(size,display,portal,latest,Encoder::Cpu,true,sequence,settings,idr,budget).await; }
#[cfg_attr(not(feature = "inproc-av"), allow(unused_mut))]
let mut result = read_encoder_with(size,display.clone(),portal.clone(),latest.clone(),encoder,true,sequence.clone(),settings,idr.clone(),budget.clone()).await;
// A failed native NVENC session (old driver, unsupported mode) still leaves FFmpeg's NVENC process pipeline.
#[cfg(feature = "inproc-av")]
if let Err(error) = &result {
if !latest.is_closed() && native_inproc(&display, portal.is_some(), encoder) {
log::warn!("In-process {} stream unavailable; switching to the FFmpeg process pipeline: {}",encoder.name_for(settings.codec),error);
result = read_encoder_with(size,display.clone(),portal.clone(),latest.clone(),encoder,false,sequence.clone(),settings,idr.clone(),budget.clone()).await;
}
}
match result {
Ok(())=>Ok(()),
Err(error) if !latest.is_closed()=>{
log::warn!("{} stream unavailable; switching scoped encoder to CPU: {}",encoder.name_for(settings.codec),error);
read_encoder_with(size,display,portal,latest,Encoder::Cpu,sequence,settings,idr,budget).await
read_encoder_with(size,display,portal,latest,Encoder::Cpu,true,sequence,settings,idr,budget).await
},
Err(error)=>Err(error),
}
}
/// The native in-process NVENC path serves X11 only (not Wayland/KMS capture or a portal session).
#[cfg(feature = "inproc-av")]
fn native_inproc(display: &str, portal: bool, encoder: Encoder) -> bool {
!display.starts_with("wayland:") && !display.starts_with("kms:") && !portal && crate::inproc_encoder::selected(encoder)
}
/// Capture and encode inside this process; each packet is one access unit.
#[cfg(feature = "inproc-av")]
async fn read_inprocess(
@@ -528,9 +542,11 @@ async fn read_inprocess(
let stop = Arc::new(AtomicBool::new(false));
// If this future is dropped (reader task aborted), the blocking encoder stops at its next packet.
let _stop_on_drop = StopOnDrop(stop.clone());
// A restart aborts the previous reader without waiting for its blocking encoder; wait for it here.
let slot = crate::inproc_encoder::stream_slot(&idr).await;
let job = tokio::task::spawn_blocking({
let stop = stop.clone();
move || crate::inproc_encoder::run(config, units_tx, stop, idr, budget)
move || { let _slot = slot; crate::inproc_encoder::run(config, units_tx, stop, idr, budget) }
});
let mut gate = DependencyGate::default();
while let Some(unit) = units.recv().await {
@@ -579,7 +595,8 @@ fn rate_args(settings: &VideoSettings, encoder: Encoder) -> (String, String) {
}
async fn read_encoder_with(
size: (u32,u32),display:String,portal:Option<Arc<crate::portal::PortalSession>>,latest:mpsc::Sender<Arc<CompleteFrame>>,encoder:Encoder,sequence:Arc<std::sync::atomic::AtomicU64>,
size: (u32,u32),display:String,portal:Option<Arc<crate::portal::PortalSession>>,latest:mpsc::Sender<Arc<CompleteFrame>>,encoder:Encoder,
#[cfg_attr(not(feature = "inproc-av"), allow(unused_variables))] allow_inproc: bool,sequence:Arc<std::sync::atomic::AtomicU64>,
settings: VideoSettings, #[cfg_attr(not(feature = "inproc-av"), allow(unused_variables))] idr: Arc<AtomicBool>, budget: Arc<Semaphore>,
)->Result<(),String>{
let selected_output=if display.starts_with("wayland:") {crate::wayland::source(settings.source_id)} else {None};
@@ -593,14 +610,14 @@ async fn read_encoder_with(
Some(Arc::new(capture))
} else { None };
#[cfg(feature = "inproc-av")]
if !wayland && portal.is_none() && crate::inproc_encoder::selected() {
if allow_inproc && native_inproc(&display, portal.is_some(), encoder) {
let max_kbps = settings.max_kbps().unwrap_or(30_000) as i64;
let config = crate::inproc_encoder::Config {
display: display.clone(), input: size, output: settings.output, fps: u32::from(settings.fps),
gop: gop_frames(settings.fps, true) as u32, cq: settings.cq() as u32,
target_bps: settings.max_kbps().map_or(16_000, |k| i64::from(k) * 3 / 4) * 1000,
max_bps: max_kbps * 1000, buffer_bits: i64::from(buffer_for(&settings, encoder)) * 1000, encoder, codec: settings.codec,
damage: crate::inproc_encoder::damage_enabled(),
max_bps: max_kbps * 1000, buffer_bits: i64::from(buffer_for(&settings, encoder)) * 1000, codec: settings.codec, chroma: settings.chroma,
vaapi: encoder == Encoder::Vaapi,
};
log::info!("Desktop encoder: display={} source={}x{} output={:?} fps={} cq={} maxrate={}k encoder={} pipeline=in-process",
display,size.0,size.1,settings.output,settings.fps,config.cq,max_kbps,encoder.name_for(settings.codec));
@@ -775,7 +792,10 @@ fn probe_settings() -> Result<VideoSettings,String> {
None=>crate::capture_setting::QUALITY_AUTO,
Some(v)=>v.parse::<u64>().ok().filter(|q|(1..=5).contains(q)).ok_or("UU_PROBE_QUALITY must be 1-5")?,
};
Ok(VideoSettings {fps,quality,codec:probe_codec(),..VideoSettings::default()})
let mut settings=VideoSettings {fps,quality,codec:probe_codec(),..VideoSettings::default()};
// UU_PROBE_CHROMA=444 requests 真彩 like the viewer does (only where full_chroma_available()).
if std::env::var("UU_PROBE_CHROMA").as_deref()==Ok("444") && full_chroma_available() { settings.chroma=crate::capture_setting::CHROMA_444; }
Ok(settings)
}
/// The probe follows `UU_VIDEO_CODEC` so the HEVC path can be measured at all (H.264 otherwise),
+1 -1
View File
@@ -1,4 +1,4 @@
//! Uses the same optional FFmpeg dependency as the host inproc-av feature.
//! Uses the optional FFmpeg 6 libraries of the `viewer-ffmpeg` feature.
use super::desktop::{Decoder,Frame};
use ffmpeg_next as av;
use super::media::Codec;
+4
View File
@@ -0,0 +1,4 @@
`ffnvcodec/dynlink_cuda.h` and `ffnvcodec/nvEncodeAPI.h` are unmodified copies from
[nv-codec-headers](https://github.com/FFmpeg/nv-codec-headers) tag `n12.1.14.0`
(NVENC API 12.1, the version FFmpeg 6.1 builds against; NVIDIA Linux driver 530.41 or newer).
MIT licensed, see the header comments. Only the in-process NVENC engine (`--features inproc-av`) uses them.
+518
View File
@@ -0,0 +1,518 @@
/*
* This copyright notice applies to this header file only:
*
* Copyright (c) 2016
*
* Permission is hereby granted, free of charge, to any person
* obtaining a copy of this software and associated documentation
* files (the "Software"), to deal in the Software without
* restriction, including without limitation the rights to use,
* copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the software, and to permit persons to whom the
* software is furnished to do so, subject to the following
* conditions:
*
* The above copyright notice and this permission notice shall be
* included in all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
* OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
* HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
* WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
* OTHER DEALINGS IN THE SOFTWARE.
*/
#if !defined(FFNV_DYNLINK_CUDA_H) && !defined(CUDA_VERSION)
#define FFNV_DYNLINK_CUDA_H
#include <stddef.h>
#include <stdint.h>
#define CUDA_VERSION 7050
#if defined(_WIN32) || defined(__CYGWIN__)
#define CUDAAPI __stdcall
#else
#define CUDAAPI
#endif
#define CU_CTX_SCHED_BLOCKING_SYNC 4
typedef int CUdevice;
#if defined(__x86_64) || defined(AMD64) || defined(_M_AMD64) || defined(__LP64__) || defined(__aarch64__)
typedef unsigned long long CUdeviceptr;
#else
typedef unsigned int CUdeviceptr;
#endif
typedef unsigned long long CUtexObject;
typedef struct CUarray_st *CUarray;
typedef struct CUctx_st *CUcontext;
typedef struct CUstream_st *CUstream;
typedef struct CUevent_st *CUevent;
typedef struct CUfunc_st *CUfunction;
typedef struct CUmod_st *CUmodule;
typedef struct CUmipmappedArray_st *CUmipmappedArray;
typedef struct CUgraphicsResource_st *CUgraphicsResource;
typedef struct CUextMemory_st *CUexternalMemory;
typedef struct CUextSemaphore_st *CUexternalSemaphore;
typedef struct CUeglStreamConnection_st *CUeglStreamConnection;
typedef struct CUlinkState_st *CUlinkState;
typedef enum cudaError_enum {
CUDA_SUCCESS = 0,
CUDA_ERROR_NOT_READY = 600,
CUDA_ERROR_LAUNCH_TIMEOUT = 702,
CUDA_ERROR_UNKNOWN = 999
} CUresult;
/**
* Device properties (subset)
*/
typedef enum CUdevice_attribute_enum {
CU_DEVICE_ATTRIBUTE_CLOCK_RATE = 13,
CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT = 14,
CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT = 16,
CU_DEVICE_ATTRIBUTE_INTEGRATED = 18,
CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY = 19,
CU_DEVICE_ATTRIBUTE_COMPUTE_MODE = 20,
CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS = 31,
CU_DEVICE_ATTRIBUTE_PCI_BUS_ID = 33,
CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID = 34,
CU_DEVICE_ATTRIBUTE_TCC_DRIVER = 35,
CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE = 36,
CU_DEVICE_ATTRIBUTE_GLOBAL_MEMORY_BUS_WIDTH = 37,
CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT = 40,
CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING = 41,
CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID = 50,
CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT = 51,
CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR = 75,
CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR = 76,
CU_DEVICE_ATTRIBUTE_MANAGED_MEMORY = 83,
CU_DEVICE_ATTRIBUTE_MULTI_GPU_BOARD = 84,
CU_DEVICE_ATTRIBUTE_MULTI_GPU_BOARD_GROUP_ID = 85,
} CUdevice_attribute;
typedef enum CUarray_format_enum {
CU_AD_FORMAT_UNSIGNED_INT8 = 0x01,
CU_AD_FORMAT_UNSIGNED_INT16 = 0x02,
CU_AD_FORMAT_UNSIGNED_INT32 = 0x03,
CU_AD_FORMAT_SIGNED_INT8 = 0x08,
CU_AD_FORMAT_SIGNED_INT16 = 0x09,
CU_AD_FORMAT_SIGNED_INT32 = 0x0a,
CU_AD_FORMAT_HALF = 0x10,
CU_AD_FORMAT_FLOAT = 0x20
} CUarray_format;
typedef enum CUmemorytype_enum {
CU_MEMORYTYPE_HOST = 1,
CU_MEMORYTYPE_DEVICE = 2,
CU_MEMORYTYPE_ARRAY = 3
} CUmemorytype;
typedef enum CUlimit_enum {
CU_LIMIT_STACK_SIZE = 0,
CU_LIMIT_PRINTF_FIFO_SIZE = 1,
CU_LIMIT_MALLOC_HEAP_SIZE = 2,
CU_LIMIT_DEV_RUNTIME_SYNC_DEPTH = 3,
CU_LIMIT_DEV_RUNTIME_PENDING_LAUNCH_COUNT = 4
} CUlimit;
typedef enum CUresourcetype_enum {
CU_RESOURCE_TYPE_ARRAY = 0x00,
CU_RESOURCE_TYPE_MIPMAPPED_ARRAY = 0x01,
CU_RESOURCE_TYPE_LINEAR = 0x02,
CU_RESOURCE_TYPE_PITCH2D = 0x03
} CUresourcetype;
typedef enum CUaddress_mode_enum {
CU_TR_ADDRESS_MODE_WRAP = 0,
CU_TR_ADDRESS_MODE_CLAMP = 1,
CU_TR_ADDRESS_MODE_MIRROR = 2,
CU_TR_ADDRESS_MODE_BORDER = 3
} CUaddress_mode;
typedef enum CUfilter_mode_enum {
CU_TR_FILTER_MODE_POINT = 0,
CU_TR_FILTER_MODE_LINEAR = 1
} CUfilter_mode;
typedef enum CUgraphicsRegisterFlags_enum {
CU_GRAPHICS_REGISTER_FLAGS_NONE = 0,
CU_GRAPHICS_REGISTER_FLAGS_READ_ONLY = 1,
CU_GRAPHICS_REGISTER_FLAGS_WRITE_DISCARD = 2,
CU_GRAPHICS_REGISTER_FLAGS_SURFACE_LDST = 4,
CU_GRAPHICS_REGISTER_FLAGS_TEXTURE_GATHER = 8
} CUgraphicsRegisterFlags;
typedef enum CUexternalMemoryHandleType_enum {
CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD = 1,
CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32 = 2,
CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT = 3,
CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP = 4,
CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE = 5,
} CUexternalMemoryHandleType;
typedef enum CUexternalSemaphoreHandleType_enum {
CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD = 1,
CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32 = 2,
CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT = 3,
CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE = 4,
CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_FD = 9,
CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_WIN32 = 10,
} CUexternalSemaphoreHandleType;
typedef enum CUjit_option_enum
{
CU_JIT_MAX_REGISTERS = 0,
CU_JIT_THREADS_PER_BLOCK = 1,
CU_JIT_WALL_TIME = 2,
CU_JIT_INFO_LOG_BUFFER = 3,
CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES = 4,
CU_JIT_ERROR_LOG_BUFFER = 5,
CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES = 6,
CU_JIT_OPTIMIZATION_LEVEL = 7,
CU_JIT_TARGET_FROM_CUCONTEXT = 8,
CU_JIT_TARGET = 9,
CU_JIT_FALLBACK_STRATEGY = 10,
CU_JIT_GENERATE_DEBUG_INFO = 11,
CU_JIT_LOG_VERBOSE = 12,
CU_JIT_GENERATE_LINE_INFO = 13,
CU_JIT_CACHE_MODE = 14,
CU_JIT_NEW_SM3X_OPT = 15,
CU_JIT_FAST_COMPILE = 16,
CU_JIT_GLOBAL_SYMBOL_NAMES = 17,
CU_JIT_GLOBAL_SYMBOL_ADDRESSES = 18,
CU_JIT_GLOBAL_SYMBOL_COUNT = 19,
CU_JIT_NUM_OPTIONS
} CUjit_option;
typedef enum CUjitInputType_enum
{
CU_JIT_INPUT_CUBIN = 0,
CU_JIT_INPUT_PTX = 1,
CU_JIT_INPUT_FATBINARY = 2,
CU_JIT_INPUT_OBJECT = 3,
CU_JIT_INPUT_LIBRARY = 4,
CU_JIT_NUM_INPUT_TYPES
} CUjitInputType;
typedef enum CUeglFrameType
{
CU_EGL_FRAME_TYPE_ARRAY = 0,
CU_EGL_FRAME_TYPE_PITCH = 1,
} CUeglFrameType;
typedef enum CUeglColorFormat
{
CU_EGL_COLOR_FORMAT_YUV420_PLANAR = 0x00,
CU_EGL_COLOR_FORMAT_YUV420_SEMIPLANAR = 0x01,
CU_EGL_COLOR_FORMAT_YVU420_SEMIPLANAR = 0x15,
CU_EGL_COLOR_FORMAT_Y10V10U10_420_SEMIPLANAR = 0x17,
CU_EGL_COLOR_FORMAT_Y12V12U12_420_SEMIPLANAR = 0x19,
} CUeglColorFormat;
typedef enum CUd3d11DeviceList_enum
{
CU_D3D11_DEVICE_LIST_ALL = 1,
CU_D3D11_DEVICE_LIST_CURRENT_FRAME = 2,
CU_D3D11_DEVICE_LIST_NEXT_FRAME = 3,
} CUd3d11DeviceList;
#ifndef CU_UUID_HAS_BEEN_DEFINED
#define CU_UUID_HAS_BEEN_DEFINED
typedef struct CUuuid_st {
char bytes[16];
} CUuuid;
#endif
typedef struct CUDA_MEMCPY2D_st {
size_t srcXInBytes;
size_t srcY;
CUmemorytype srcMemoryType;
const void *srcHost;
CUdeviceptr srcDevice;
CUarray srcArray;
size_t srcPitch;
size_t dstXInBytes;
size_t dstY;
CUmemorytype dstMemoryType;
void *dstHost;
CUdeviceptr dstDevice;
CUarray dstArray;
size_t dstPitch;
size_t WidthInBytes;
size_t Height;
} CUDA_MEMCPY2D;
typedef struct CUDA_RESOURCE_DESC_st {
CUresourcetype resType;
union {
struct {
CUarray hArray;
} array;
struct {
CUmipmappedArray hMipmappedArray;
} mipmap;
struct {
CUdeviceptr devPtr;
CUarray_format format;
unsigned int numChannels;
size_t sizeInBytes;
} linear;
struct {
CUdeviceptr devPtr;
CUarray_format format;
unsigned int numChannels;
size_t width;
size_t height;
size_t pitchInBytes;
} pitch2D;
struct {
int reserved[32];
} reserved;
} res;
unsigned int flags;
} CUDA_RESOURCE_DESC;
typedef struct CUDA_TEXTURE_DESC_st {
CUaddress_mode addressMode[3];
CUfilter_mode filterMode;
unsigned int flags;
unsigned int maxAnisotropy;
CUfilter_mode mipmapFilterMode;
float mipmapLevelBias;
float minMipmapLevelClamp;
float maxMipmapLevelClamp;
float borderColor[4];
int reserved[12];
} CUDA_TEXTURE_DESC;
/* Unused type */
typedef struct CUDA_RESOURCE_VIEW_DESC_st CUDA_RESOURCE_VIEW_DESC;
typedef unsigned int GLenum;
typedef unsigned int GLuint;
/*
* Prefix type name to avoid collisions. Clients using these types
* will include the real headers with real definitions.
*/
typedef int32_t ffnv_EGLint;
typedef void *ffnv_EGLStreamKHR;
typedef enum CUGLDeviceList_enum {
CU_GL_DEVICE_LIST_ALL = 1,
CU_GL_DEVICE_LIST_CURRENT_FRAME = 2,
CU_GL_DEVICE_LIST_NEXT_FRAME = 3,
} CUGLDeviceList;
typedef struct CUDA_EXTERNAL_MEMORY_HANDLE_DESC_st {
CUexternalMemoryHandleType type;
union {
int fd;
struct {
void *handle;
const void *name;
} win32;
} handle;
unsigned long long size;
unsigned int flags;
unsigned int reserved[16];
} CUDA_EXTERNAL_MEMORY_HANDLE_DESC;
typedef struct CUDA_EXTERNAL_MEMORY_BUFFER_DESC_st {
unsigned long long offset;
unsigned long long size;
unsigned int flags;
unsigned int reserved[16];
} CUDA_EXTERNAL_MEMORY_BUFFER_DESC;
typedef struct CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC_st {
CUexternalSemaphoreHandleType type;
union {
int fd;
struct {
void *handle;
const void *name;
} win32;
} handle;
unsigned int flags;
unsigned int reserved[16];
} CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC;
typedef struct CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS_st {
struct {
struct {
unsigned long long value;
} fence;
unsigned int reserved[16];
} params;
unsigned int flags;
unsigned int reserved[16];
} CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS;
typedef CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS;
typedef struct CUDA_ARRAY_DESCRIPTOR_st {
size_t Width;
size_t Height;
CUarray_format Format;
unsigned int NumChannels;
} CUDA_ARRAY_DESCRIPTOR;
typedef struct CUDA_ARRAY3D_DESCRIPTOR_st {
size_t Width;
size_t Height;
size_t Depth;
CUarray_format Format;
unsigned int NumChannels;
unsigned int Flags;
} CUDA_ARRAY3D_DESCRIPTOR;
typedef struct CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC_st {
unsigned long long offset;
CUDA_ARRAY3D_DESCRIPTOR arrayDesc;
unsigned int numLevels;
unsigned int reserved[16];
} CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC;
#define CU_EGL_FRAME_MAX_PLANES 3
typedef struct CUeglFrame_st {
union {
CUarray pArray[CU_EGL_FRAME_MAX_PLANES];
void* pPitch[CU_EGL_FRAME_MAX_PLANES];
} frame;
unsigned int width;
unsigned int height;
unsigned int depth;
unsigned int pitch;
unsigned int planeCount;
unsigned int numChannels;
CUeglFrameType frameType;
CUeglColorFormat eglColorFormat;
CUarray_format cuFormat;
} CUeglFrame;
#define CU_STREAM_DEFAULT 0
#define CU_STREAM_NON_BLOCKING 1
#define CU_EVENT_DEFAULT 0
#define CU_EVENT_BLOCKING_SYNC 1
#define CU_EVENT_DISABLE_TIMING 2
#define CU_EVENT_WAIT_DEFAULT 0
#define CU_EVENT_WAIT_EXTERNAL 1
#define CU_TRSF_READ_AS_INTEGER 1
typedef void CUDAAPI CUstreamCallback(CUstream hStream, CUresult status, void *userdata);
typedef CUresult CUDAAPI tcuInit(unsigned int Flags);
typedef CUresult CUDAAPI tcuDriverGetVersion(int *driverVersion);
typedef CUresult CUDAAPI tcuDeviceGetCount(int *count);
typedef CUresult CUDAAPI tcuDeviceGet(CUdevice *device, int ordinal);
typedef CUresult CUDAAPI tcuDeviceGetAttribute(int *pi, CUdevice_attribute attrib, CUdevice dev);
typedef CUresult CUDAAPI tcuDeviceGetName(char *name, int len, CUdevice dev);
typedef CUresult CUDAAPI tcuDeviceGetUuid(CUuuid *uuid, CUdevice dev);
typedef CUresult CUDAAPI tcuDeviceGetUuid_v2(CUuuid *uuid, CUdevice dev);
typedef CUresult CUDAAPI tcuDeviceGetLuid(char* luid, unsigned int* deviceNodeMask, CUdevice dev);
typedef CUresult CUDAAPI tcuDeviceGetByPCIBusId(CUdevice* dev, const char* pciBusId);
typedef CUresult CUDAAPI tcuDeviceGetPCIBusId(char* pciBusId, int len, CUdevice dev);
typedef CUresult CUDAAPI tcuDeviceComputeCapability(int *major, int *minor, CUdevice dev);
typedef CUresult CUDAAPI tcuCtxCreate_v2(CUcontext *pctx, unsigned int flags, CUdevice dev);
typedef CUresult CUDAAPI tcuCtxGetCurrent(CUcontext *pctx);
typedef CUresult CUDAAPI tcuCtxSetLimit(CUlimit limit, size_t value);
typedef CUresult CUDAAPI tcuCtxPushCurrent_v2(CUcontext pctx);
typedef CUresult CUDAAPI tcuCtxPopCurrent_v2(CUcontext *pctx);
typedef CUresult CUDAAPI tcuCtxDestroy_v2(CUcontext ctx);
typedef CUresult CUDAAPI tcuMemAlloc_v2(CUdeviceptr *dptr, size_t bytesize);
typedef CUresult CUDAAPI tcuMemAllocPitch_v2(CUdeviceptr *dptr, size_t *pPitch, size_t WidthInBytes, size_t Height, unsigned int ElementSizeBytes);
typedef CUresult CUDAAPI tcuMemAllocManaged(CUdeviceptr *dptr, size_t bytesize, unsigned int flags);
typedef CUresult CUDAAPI tcuMemsetD8Async(CUdeviceptr dstDevice, unsigned char uc, size_t N, CUstream hStream);
typedef CUresult CUDAAPI tcuMemFree_v2(CUdeviceptr dptr);
typedef CUresult CUDAAPI tcuMemcpy(CUdeviceptr dst, CUdeviceptr src, size_t bytesize);
typedef CUresult CUDAAPI tcuMemcpyAsync(CUdeviceptr dst, CUdeviceptr src, size_t bytesize, CUstream hStream);
typedef CUresult CUDAAPI tcuMemcpy2D_v2(const CUDA_MEMCPY2D *pcopy);
typedef CUresult CUDAAPI tcuMemcpy2DAsync_v2(const CUDA_MEMCPY2D *pcopy, CUstream hStream);
typedef CUresult CUDAAPI tcuMemcpyHtoD_v2(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount);
typedef CUresult CUDAAPI tcuMemcpyHtoDAsync_v2(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount, CUstream hStream);
typedef CUresult CUDAAPI tcuMemcpyDtoH_v2(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount);
typedef CUresult CUDAAPI tcuMemcpyDtoHAsync_v2(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount, CUstream hStream);
typedef CUresult CUDAAPI tcuMemcpyDtoD_v2(CUdeviceptr dstDevice, CUdeviceptr srcDevice, size_t ByteCount);
typedef CUresult CUDAAPI tcuMemcpyDtoDAsync_v2(CUdeviceptr dstDevice, CUdeviceptr srcDevice, size_t ByteCount, CUstream hStream);
typedef CUresult CUDAAPI tcuGetErrorName(CUresult error, const char** pstr);
typedef CUresult CUDAAPI tcuGetErrorString(CUresult error, const char** pstr);
typedef CUresult CUDAAPI tcuCtxGetDevice(CUdevice *device);
typedef CUresult CUDAAPI tcuDevicePrimaryCtxRetain(CUcontext *pctx, CUdevice dev);
typedef CUresult CUDAAPI tcuDevicePrimaryCtxRelease(CUdevice dev);
typedef CUresult CUDAAPI tcuDevicePrimaryCtxSetFlags(CUdevice dev, unsigned int flags);
typedef CUresult CUDAAPI tcuDevicePrimaryCtxGetState(CUdevice dev, unsigned int *flags, int *active);
typedef CUresult CUDAAPI tcuDevicePrimaryCtxReset(CUdevice dev);
typedef CUresult CUDAAPI tcuStreamCreate(CUstream *phStream, unsigned int flags);
typedef CUresult CUDAAPI tcuStreamQuery(CUstream hStream);
typedef CUresult CUDAAPI tcuStreamSynchronize(CUstream hStream);
typedef CUresult CUDAAPI tcuStreamDestroy_v2(CUstream hStream);
typedef CUresult CUDAAPI tcuStreamAddCallback(CUstream hStream, CUstreamCallback *callback, void *userdata, unsigned int flags);
typedef CUresult CUDAAPI tcuStreamWaitEvent(CUstream hStream, CUevent hEvent, unsigned int flags);
typedef CUresult CUDAAPI tcuEventCreate(CUevent *phEvent, unsigned int flags);
typedef CUresult CUDAAPI tcuEventDestroy_v2(CUevent hEvent);
typedef CUresult CUDAAPI tcuEventSynchronize(CUevent hEvent);
typedef CUresult CUDAAPI tcuEventQuery(CUevent hEvent);
typedef CUresult CUDAAPI tcuEventRecord(CUevent hEvent, CUstream hStream);
typedef CUresult CUDAAPI tcuLaunchKernel(CUfunction f, unsigned int gridDimX, unsigned int gridDimY, unsigned int gridDimZ, unsigned int blockDimX, unsigned int blockDimY, unsigned int blockDimZ, unsigned int sharedMemBytes, CUstream hStream, void** kernelParams, void** extra);
typedef CUresult CUDAAPI tcuLinkCreate(unsigned int numOptions, CUjit_option* options, void** optionValues, CUlinkState* stateOut);
typedef CUresult CUDAAPI tcuLinkAddData(CUlinkState state, CUjitInputType type, void* data, size_t size, const char* name, unsigned int numOptions, CUjit_option* options, void** optionValues);
typedef CUresult CUDAAPI tcuLinkComplete(CUlinkState state, void** cubinOut, size_t* sizeOut);
typedef CUresult CUDAAPI tcuLinkDestroy(CUlinkState state);
typedef CUresult CUDAAPI tcuModuleLoadData(CUmodule* module, const void* image);
typedef CUresult CUDAAPI tcuModuleUnload(CUmodule hmod);
typedef CUresult CUDAAPI tcuModuleGetFunction(CUfunction* hfunc, CUmodule hmod, const char* name);
typedef CUresult CUDAAPI tcuModuleGetGlobal(CUdeviceptr *dptr, size_t *bytes, CUmodule hmod, const char* name);
typedef CUresult CUDAAPI tcuTexObjectCreate(CUtexObject* pTexObject, const CUDA_RESOURCE_DESC* pResDesc, const CUDA_TEXTURE_DESC* pTexDesc, const CUDA_RESOURCE_VIEW_DESC* pResViewDesc);
typedef CUresult CUDAAPI tcuTexObjectDestroy(CUtexObject texObject);
typedef CUresult CUDAAPI tcuGLGetDevices_v2(unsigned int* pCudaDeviceCount, CUdevice* pCudaDevices, unsigned int cudaDeviceCount, CUGLDeviceList deviceList);
typedef CUresult CUDAAPI tcuGraphicsGLRegisterImage(CUgraphicsResource* pCudaResource, GLuint image, GLenum target, unsigned int Flags);
typedef CUresult CUDAAPI tcuGraphicsUnregisterResource(CUgraphicsResource resource);
typedef CUresult CUDAAPI tcuGraphicsMapResources(unsigned int count, CUgraphicsResource* resources, CUstream hStream);
typedef CUresult CUDAAPI tcuGraphicsUnmapResources(unsigned int count, CUgraphicsResource* resources, CUstream hStream);
typedef CUresult CUDAAPI tcuGraphicsSubResourceGetMappedArray(CUarray* pArray, CUgraphicsResource resource, unsigned int arrayIndex, unsigned int mipLevel);
typedef CUresult CUDAAPI tcuGraphicsResourceGetMappedPointer(CUdeviceptr *devPtrOut, size_t *sizeOut, CUgraphicsResource resource);
typedef CUresult CUDAAPI tcuImportExternalMemory(CUexternalMemory* extMem_out, const CUDA_EXTERNAL_MEMORY_HANDLE_DESC* memHandleDesc);
typedef CUresult CUDAAPI tcuDestroyExternalMemory(CUexternalMemory extMem);
typedef CUresult CUDAAPI tcuExternalMemoryGetMappedBuffer(CUdeviceptr* devPtr, CUexternalMemory extMem, const CUDA_EXTERNAL_MEMORY_BUFFER_DESC* bufferDesc);
typedef CUresult CUDAAPI tcuExternalMemoryGetMappedMipmappedArray(CUmipmappedArray* mipmap, CUexternalMemory extMem, const CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC* mipmapDesc);
typedef CUresult CUDAAPI tcuMipmappedArrayGetLevel(CUarray* pLevelArray, CUmipmappedArray hMipmappedArray, unsigned int level);
typedef CUresult CUDAAPI tcuMipmappedArrayDestroy(CUmipmappedArray hMipmappedArray);
typedef CUresult CUDAAPI tcuImportExternalSemaphore(CUexternalSemaphore* extSem_out, const CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC* semHandleDesc);
typedef CUresult CUDAAPI tcuDestroyExternalSemaphore(CUexternalSemaphore extSem);
typedef CUresult CUDAAPI tcuSignalExternalSemaphoresAsync(const CUexternalSemaphore* extSemArray, const CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS* paramsArray, unsigned int numExtSems, CUstream stream);
typedef CUresult CUDAAPI tcuWaitExternalSemaphoresAsync(const CUexternalSemaphore* extSemArray, const CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS* paramsArray, unsigned int numExtSems, CUstream stream);
typedef CUresult CUDAAPI tcuArrayCreate(CUarray *pHandle, const CUDA_ARRAY_DESCRIPTOR* pAllocateArray);
typedef CUresult CUDAAPI tcuArray3DCreate(CUarray *pHandle, const CUDA_ARRAY3D_DESCRIPTOR* pAllocateArray);
typedef CUresult CUDAAPI tcuArrayDestroy(CUarray hArray);
typedef CUresult CUDAAPI tcuEGLStreamProducerConnect(CUeglStreamConnection* conn, ffnv_EGLStreamKHR stream, ffnv_EGLint width, ffnv_EGLint height);
typedef CUresult CUDAAPI tcuEGLStreamProducerDisconnect(CUeglStreamConnection* conn);
typedef CUresult CUDAAPI tcuEGLStreamConsumerDisconnect(CUeglStreamConnection* conn);
typedef CUresult CUDAAPI tcuEGLStreamProducerPresentFrame(CUeglStreamConnection* conn, CUeglFrame eglframe, CUstream* pStream);
typedef CUresult CUDAAPI tcuEGLStreamProducerReturnFrame(CUeglStreamConnection* conn, CUeglFrame* eglframe, CUstream* pStream);
typedef CUresult CUDAAPI tcuD3D11GetDevice(CUdevice *device, void *dxgiAdapter);
typedef CUresult CUDAAPI tcuD3D11GetDevices(unsigned int *deviceCountOut, CUdevice *devices, unsigned int deviceCount, void *d3d11device, CUd3d11DeviceList listType);
typedef CUresult CUDAAPI tcuGraphicsD3D11RegisterResource(CUgraphicsResource *cudaResourceOut, void *d3d11Resource, unsigned int flags);
#endif
File diff suppressed because it is too large. Load diff