perf(host/windows): move capture→encode off the 3D engine (NV12/P010 video-processor path, zero-copy, GPU priority)

The Windows host capped at ~60 fps with 35-40 ms latency on a GPU-heavy game: the per-frame capture→encode path shared the 3D engine with the game and got scheduled behind it. Rework to minimize 3D-engine work per frame: - VideoConverter (D3D11 video processor): capture → NVENC-native NV12/P010 so NVENC skips its internal RGB→YUV (a 3D/compute step). Wired into both DDA (dxgi.rs) and WGC (wgc.rs). New PixelFormat::Nv12/P010 + NVENC YUV input. - GPU scheduling hardening (Apollo-style): D3DKMTSetProcessSchedulingPriorityClass HIGH, absolute SetGPUThreadPriority, SetMaximumFrameLatency(1). - WGC SDR zero-copy (hold pool frames; no CopyResource). DDA keeps a fast CopyResource to decouple its single-frame acquire/release from the async convert. - Pipelined helper encode loop (PUNKTFUNK_ENCODE_DEPTH, default 1) + perf split (cap_wait / encode / write). Live on the RTX 4090: hard 60 fps ceiling removed (now scene-scaling 40-200+), latency much reduced. Residual cap in GPU-pinned scenes is the irreducible RGB→YUV convert (no fixed-function unit on NVIDIA — VideoProcessing engine reads 0%) waiting behind an uncapped game under WDDM context time-slicing; Linux avoids it via gamescope capping the game to the display refresh. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-17 13:08:03 +00:00
parent 15d3d423fa
commit 4cc57d5c39
8 changed files with 689 additions and 67 deletions
@@ -26,6 +26,13 @@ pub enum PixelFormat {
    /// produces this: scRGB FP16 desktop pixels are converted to BT.2020 PQ and written here, then
    /// handed to NVENC as `ABGR10` for an HEVC Main10 / HDR10 encode.
    Rgb10a2,
+    /// `NV12` (DXGI `NV12`): 8-bit BT.709 limited-range YUV 4:2:0. Produced by the D3D11 **video
+    /// processor** (video engine, not the 3D engine) so the per-frame colour conversion doesn't fight a
+    /// GPU-saturating game; handed to NVENC as `NV12` (it encodes YUV natively — no internal RGB→YUV).
+    Nv12,
+    /// `P010` (DXGI `P010`): 10-bit BT.2020 PQ limited-range YUV 4:2:0. HDR analogue of [`Nv12`]:
+    /// video-processor output for HEVC Main10 / HDR10, handed to NVENC as `YUV420_10BIT`.
+    P010,
 }

 impl PixelFormat {
@@ -37,12 +37,12 @@ use windows::Win32::Graphics::Dxgi::Common::{
    DXGI_FORMAT_R16G16B16A16_FLOAT, DXGI_SAMPLE_DESC,
 };
 use windows::Win32::Graphics::Dxgi::{
-    CreateDXGIFactory1, IDXGIAdapter1, IDXGIFactory1, IDXGIOutput1, IDXGIOutput5,
-    IDXGIOutputDuplication, IDXGIResource, DXGI_ERROR_ACCESS_LOST, DXGI_ERROR_DEVICE_REMOVED,
-    DXGI_ERROR_DEVICE_RESET, DXGI_ERROR_INVALID_CALL, DXGI_ERROR_MODE_CHANGE_IN_PROGRESS,
-    DXGI_ERROR_WAIT_TIMEOUT, DXGI_OUTDUPL_DESC, DXGI_OUTDUPL_FRAME_INFO,
-    DXGI_OUTDUPL_POINTER_SHAPE_INFO, DXGI_OUTDUPL_POINTER_SHAPE_TYPE_COLOR,
-    DXGI_OUTDUPL_POINTER_SHAPE_TYPE_MASKED_COLOR,
+    CreateDXGIFactory1, IDXGIAdapter1, IDXGIDevice, IDXGIDevice1, IDXGIFactory1, IDXGIOutput1,
+    IDXGIOutput5, IDXGIOutputDuplication, IDXGIResource, DXGI_ERROR_ACCESS_LOST,
+    DXGI_ERROR_DEVICE_REMOVED, DXGI_ERROR_DEVICE_RESET, DXGI_ERROR_INVALID_CALL,
+    DXGI_ERROR_MODE_CHANGE_IN_PROGRESS, DXGI_ERROR_WAIT_TIMEOUT, DXGI_OUTDUPL_DESC,
+    DXGI_OUTDUPL_FRAME_INFO, DXGI_OUTDUPL_POINTER_SHAPE_INFO,
+    DXGI_OUTDUPL_POINTER_SHAPE_TYPE_COLOR, DXGI_OUTDUPL_POINTER_SHAPE_TYPE_MASKED_COLOR,
 };
 use windows::Win32::System::StationsAndDesktops::{
    CloseDesktop, OpenInputDesktop, SetThreadDesktop, DESKTOP_ACCESS_FLAGS, DESKTOP_CONTROL_FLAGS,
@@ -147,10 +147,119 @@ pub(crate) unsafe fn make_device(
        Some(&mut context),
    )
    .context("D3D11CreateDevice")?;
-    Ok((
-        device.context("null D3D11 device")?,
-        context.context("null D3D11 context")?,
-    ))
+    let device = device.context("null D3D11 device")?;
+    let context = context.context("null D3D11 context")?;
+
+    // Apollo-style GPU scheduling hardening (Sunshine display_base.cpp:599-709). Our capture+encode
+    // shares the GPU with the streamed game; when the game saturates the GPU our process is starved of
+    // GPU time slices, so NVENC sits near-idle yet `lock_bitstream` waits ~20 ms for our context to be
+    // scheduled — capping the stream (~47 fps measured at 5K@240) and stuttering. Per-frame copy/convert
+    // is NOT the cause (zero-copy + thread-priority alone didn't move it); the PROCESS-level GPU
+    // scheduling priority class is the decisive cross-process lever. Secondary: the absolute per-device
+    // GPU thread priority and a 1-frame latency cap.
+    elevate_process_gpu_priority();
+    if let Ok(dxgi_dev) = device.cast::<IDXGIDevice>() {
+        // Apollo's absolute max GPU thread priority (0x4000001E); fall back to relative +7.
+        if dxgi_dev.SetGPUThreadPriority(0x4000_001E).is_err()
+            && dxgi_dev.SetGPUThreadPriority(7).is_err()
+        {
+            tracing::warn!("SetGPUThreadPriority failed (run as admin/SYSTEM for GPU priority)");
+        }
+    }
+    if let Ok(dxgi1) = device.cast::<IDXGIDevice1>() {
+        let _ = dxgi1.SetMaximumFrameLatency(1);
+    }
+    Ok((device, context))
+}
+
+/// Apollo-style GPU scheduling-priority hardening (Sunshine `display_base.cpp:599-709`). On a
+/// GPU-saturated game our capture+encode process is starved of GPU time slices — NVENC sits ~idle but
+/// `lock_bitstream` waits ~20 ms for our context to be scheduled. Elevating the PROCESS GPU scheduling
+/// priority class (the strong cross-process lever — far more effective than `SetGPUThreadPriority`
+/// alone, which we measured as no help) lets our brief encode preempt the game. Uses HIGH, NOT
+/// realtime: realtime on NVIDIA + HAGS can freeze/crash NVENC (Apollo downgrades it for exactly this).
+/// Runs once per process; best-effort. `PUNKTFUNK_GPU_PRIORITY_CLASS = off|normal|high|realtime`
+/// (default high).
+fn elevate_process_gpu_priority() {
+    use std::sync::Once;
+    static ONCE: Once = Once::new();
+    ONCE.call_once(|| unsafe {
+        use windows::core::{s, PCWSTR};
+        use windows::Win32::Foundation::{CloseHandle, HANDLE, LUID};
+        use windows::Win32::Security::{
+            AdjustTokenPrivileges, LookupPrivilegeValueW, LUID_AND_ATTRIBUTES,
+            SE_INC_BASE_PRIORITY_NAME, SE_PRIVILEGE_ENABLED, TOKEN_ADJUST_PRIVILEGES,
+            TOKEN_PRIVILEGES, TOKEN_QUERY,
+        };
+        use windows::Win32::System::LibraryLoader::{GetProcAddress, LoadLibraryA};
+        use windows::Win32::System::Threading::{GetCurrentProcess, OpenProcessToken};
+
+        // D3DKMT_SCHEDULINGPRIORITYCLASS: IDLE 0, BELOW_NORMAL 1, NORMAL 2, ABOVE_NORMAL 3, HIGH 4,
+        // REALTIME 5.
+        let prio: i32 = match std::env::var("PUNKTFUNK_GPU_PRIORITY_CLASS").ok().as_deref() {
+            Some("off") => {
+                tracing::info!("GPU process scheduling priority class left at default (off)");
+                return;
+            }
+            Some("normal") => 2,
+            Some("realtime") => 5,
+            _ => 4, // HIGH — safe on NVIDIA+HAGS (realtime can freeze NVENC)
+        };
+
+        // 1. Enable SE_INC_BASE_PRIORITY so the kernel permits the GPU priority bump.
+        let mut token = HANDLE::default();
+        if OpenProcessToken(
+            GetCurrentProcess(),
+            TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY,
+            &mut token,
+        )
+        .is_ok()
+        {
+            let mut luid = LUID::default();
+            if LookupPrivilegeValueW(PCWSTR::null(), SE_INC_BASE_PRIORITY_NAME, &mut luid).is_ok() {
+                let tp = TOKEN_PRIVILEGES {
+                    PrivilegeCount: 1,
+                    Privileges: [LUID_AND_ATTRIBUTES {
+                        Luid: luid,
+                        Attributes: SE_PRIVILEGE_ENABLED,
+                    }],
+                };
+                if AdjustTokenPrivileges(
+                    token,
+                    false,
+                    Some(&tp as *const TOKEN_PRIVILEGES),
+                    0,
+                    None,
+                    None,
+                )
+                .is_err()
+                {
+                    tracing::warn!("could not enable SE_INC_BASE_PRIORITY for GPU priority");
+                }
+            }
+            let _ = CloseHandle(token);
+        }
+
+        // 2. D3DKMTSetProcessSchedulingPriorityClass via gdi32 (no stable windows-rs binding).
+        if let Ok(gdi32) = LoadLibraryA(s!("gdi32.dll")) {
+            if let Some(p) = GetProcAddress(gdi32, s!("D3DKMTSetProcessSchedulingPriorityClass")) {
+                type SetPrio = unsafe extern "system" fn(HANDLE, i32) -> i32;
+                let f: SetPrio = std::mem::transmute(p);
+                let st = f(GetCurrentProcess(), prio);
+                if st == 0 {
+                    tracing::info!(
+                        priority_class = prio,
+                        "GPU process scheduling priority class set (2=normal 4=high 5=realtime)"
+                    );
+                } else {
+                    tracing::warn!(
+                        status = format!("0x{st:08X}"),
+                        "D3DKMTSetProcessSchedulingPriorityClass failed (run as admin/SYSTEM for GPU priority)"
+                    );
+                }
+            }
+        }
+    });
 }

 /// Re-find the output, make a fresh device on its adapter, and duplicate it. Used by the ACCESS_LOST
@@ -827,6 +936,135 @@ impl HdrConverter {
    }
 }

+use windows::Win32::Graphics::Direct3D11::{
+    ID3D11VideoContext1, ID3D11VideoDevice, ID3D11VideoProcessor, ID3D11VideoProcessorEnumerator,
+    ID3D11VideoProcessorInputView, ID3D11VideoProcessorOutputView, D3D11_TEX2D_VPIV,
+    D3D11_TEX2D_VPOV, D3D11_VIDEO_FRAME_FORMAT_PROGRESSIVE, D3D11_VIDEO_PROCESSOR_CONTENT_DESC,
+    D3D11_VIDEO_PROCESSOR_INPUT_VIEW_DESC, D3D11_VIDEO_PROCESSOR_INPUT_VIEW_DESC_0,
+    D3D11_VIDEO_PROCESSOR_OUTPUT_VIEW_DESC, D3D11_VIDEO_PROCESSOR_OUTPUT_VIEW_DESC_0,
+    D3D11_VIDEO_PROCESSOR_STREAM, D3D11_VIDEO_USAGE_PLAYBACK_NORMAL,
+    D3D11_VPIV_DIMENSION_TEXTURE2D, D3D11_VPOV_DIMENSION_TEXTURE2D,
+};
+use windows::Win32::Graphics::Dxgi::Common::{
+    DXGI_COLOR_SPACE_RGB_FULL_G10_NONE_P709, DXGI_COLOR_SPACE_RGB_FULL_G22_NONE_P709,
+    DXGI_COLOR_SPACE_YCBCR_STUDIO_G2084_LEFT_P2020, DXGI_COLOR_SPACE_YCBCR_STUDIO_G22_LEFT_P709,
+    DXGI_RATIONAL,
+};
+
+/// D3D11 **Video Processor** colour/format converter — runs on the GPU's dedicated VIDEO engine, NOT
+/// the 3D engine, so the per-frame RGB→YUV conversion does not contend with a GPU-saturating game (the
+/// HDR pixel-shader path and NVENC's internal RGB→YUV both use the 3D/compute engine, which an AAA
+/// title pins at ~100%). Output is NV12 (SDR, BT.709 studio-range) or P010 (HDR, BT.2020 PQ
+/// studio-range) — NVENC's native YUV inputs, so it encodes them with no further conversion.
+pub(crate) struct VideoConverter {
+    vdev: ID3D11VideoDevice,
+    vctx: ID3D11VideoContext1,
+    enumr: ID3D11VideoProcessorEnumerator,
+    vp: ID3D11VideoProcessor,
+}
+
+impl VideoConverter {
+    pub(crate) unsafe fn new(
+        device: &ID3D11Device,
+        context: &ID3D11DeviceContext,
+        width: u32,
+        height: u32,
+        hdr: bool,
+    ) -> Result<Self> {
+        let vdev: ID3D11VideoDevice = device.cast().context("device -> ID3D11VideoDevice")?;
+        let vctx: ID3D11VideoContext1 = context.cast().context("context -> ID3D11VideoContext1")?;
+        let rate = DXGI_RATIONAL {
+            Numerator: 240,
+            Denominator: 1,
+        };
+        let desc = D3D11_VIDEO_PROCESSOR_CONTENT_DESC {
+            InputFrameFormat: D3D11_VIDEO_FRAME_FORMAT_PROGRESSIVE,
+            InputFrameRate: rate,
+            InputWidth: width,
+            InputHeight: height,
+            OutputFrameRate: rate,
+            OutputWidth: width,
+            OutputHeight: height,
+            Usage: D3D11_VIDEO_USAGE_PLAYBACK_NORMAL,
+        };
+        let enumr = vdev
+            .CreateVideoProcessorEnumerator(&desc)
+            .context("CreateVideoProcessorEnumerator")?;
+        let vp = vdev
+            .CreateVideoProcessor(&enumr, 0)
+            .context("CreateVideoProcessor")?;
+
+        // Full-range RGB in → studio-range YUV out. HDR: scRGB linear (G10) → BT.2020 PQ (G2084).
+        // SDR: sRGB (G22) → BT.709 (G22).
+        let (in_cs, out_cs) = if hdr {
+            (
+                DXGI_COLOR_SPACE_RGB_FULL_G10_NONE_P709,
+                DXGI_COLOR_SPACE_YCBCR_STUDIO_G2084_LEFT_P2020,
+            )
+        } else {
+            (
+                DXGI_COLOR_SPACE_RGB_FULL_G22_NONE_P709,
+                DXGI_COLOR_SPACE_YCBCR_STUDIO_G22_LEFT_P709,
+            )
+        };
+        vctx.VideoProcessorSetStreamColorSpace1(&vp, 0, in_cs);
+        vctx.VideoProcessorSetOutputColorSpace1(&vp, out_cs);
+        // One frame in, one frame out — no interpolation/auto-processing.
+        vctx.VideoProcessorSetStreamFrameFormat(&vp, 0, D3D11_VIDEO_FRAME_FORMAT_PROGRESSIVE);
+
+        Ok(Self {
+            vdev,
+            vctx,
+            enumr,
+            vp,
+        })
+    }
+
+    /// Convert `input` (BGRA or scRGB FP16) → `output` (NV12 or P010) on the video engine. Views are
+    /// created per call (cheap relative to the Blt) so the input texture can vary frame to frame.
+    pub(crate) unsafe fn convert(
+        &self,
+        input: &ID3D11Texture2D,
+        output: &ID3D11Texture2D,
+    ) -> Result<()> {
+        let in_desc = D3D11_VIDEO_PROCESSOR_INPUT_VIEW_DESC {
+            FourCC: 0,
+            ViewDimension: D3D11_VPIV_DIMENSION_TEXTURE2D,
+            Anonymous: D3D11_VIDEO_PROCESSOR_INPUT_VIEW_DESC_0 {
+                Texture2D: D3D11_TEX2D_VPIV {
+                    MipSlice: 0,
+                    ArraySlice: 0,
+                },
+            },
+        };
+        let mut in_view: Option<ID3D11VideoProcessorInputView> = None;
+        self.vdev
+            .CreateVideoProcessorInputView(input, &self.enumr, &in_desc, Some(&mut in_view))
+            .context("CreateVideoProcessorInputView")?;
+
+        let out_desc = D3D11_VIDEO_PROCESSOR_OUTPUT_VIEW_DESC {
+            ViewDimension: D3D11_VPOV_DIMENSION_TEXTURE2D,
+            Anonymous: D3D11_VIDEO_PROCESSOR_OUTPUT_VIEW_DESC_0 {
+                Texture2D: D3D11_TEX2D_VPOV { MipSlice: 0 },
+            },
+        };
+        let mut out_view: Option<ID3D11VideoProcessorOutputView> = None;
+        self.vdev
+            .CreateVideoProcessorOutputView(output, &self.enumr, &out_desc, Some(&mut out_view))
+            .context("CreateVideoProcessorOutputView")?;
+        let out_view = out_view.context("null output view")?;
+
+        let stream = D3D11_VIDEO_PROCESSOR_STREAM {
+            Enable: true.into(),
+            pInputSurface: std::mem::ManuallyDrop::new(in_view),
+            ..Default::default()
+        };
+        self.vctx
+            .VideoProcessorBlt(&self.vp, &out_view, 0, &[stream])
+            .context("VideoProcessorBlt")
+    }
+}
+
 /// Convert a DXGI pointer shape (color / masked-color / monochrome) into top-down BGRA.
 fn convert_pointer_shape(buf: &[u8], si: &DXGI_OUTDUPL_POINTER_SHAPE_INFO) -> Option<CursorShape> {
    let w = si.Width as usize;
@@ -1055,6 +1293,17 @@ pub struct DuplCapturer {
    hdr10_out: Option<ID3D11Texture2D>,
    /// scRGB→PQ conversion pass; rebuilt on device recreate.
    hdr_conv: Option<HdrConverter>,
+    /// Video-processor RGB→YUV converter (runs on the VIDEO engine, not the 3D engine) + its NV12
+    /// (SDR) / P010 (HDR) output texture. This is the zero-3D path: the per-frame colour conversion and
+    /// NVENC's RGB→YUV both move off the 3D engine so capture+encode don't fight a GPU-saturating game.
+    /// Lazily built for the current size+HDR; rebuilt on change. `None`/error → falls back to the
+    /// legacy RGB path. Disabled with `PUNKTFUNK_NO_VIDEO_PROCESSOR=1`.
+    video_conv: Option<VideoConverter>,
+    yuv_out: Option<ID3D11Texture2D>,
+    /// HDR-ness the current `video_conv`/`yuv_out` were built for, so an HDR toggle rebuilds them.
+    yuv_is_hdr: bool,
+    /// Latched off after a VideoConverter failure so we don't retry it every frame (fall back to RGB).
+    vp_disabled: bool,
    /// Last time a duplication rebuild was attempted, to throttle retries during an outage (e.g. a
    /// secure-desktop dwell where the output is gone) so we don't block the encode loop or hammer
    /// DuplicateOutput — between attempts the last good frame is repeated. `None` = never attempted.
@@ -1306,6 +1555,10 @@ impl DuplCapturer {
                fp16_srv: None,
                hdr10_out: None,
                hdr_conv: None,
+                video_conv: None,
+                yuv_out: None,
+                yuv_is_hdr: false,
+                vp_disabled: std::env::var_os("PUNKTFUNK_NO_VIDEO_PROCESSOR").is_some(),
                last_rebuild: None,
                last_recover: None,
                ever_got_frame: false,
@@ -1375,6 +1628,85 @@ impl DuplCapturer {
        Ok(())
    }

+    /// Convert `input` (BGRA for SDR, scRGB FP16 for HDR) to NVENC's native YUV (NV12 / P010) via the
+    /// D3D11 **video processor** (video engine) — keeping the per-frame colour conversion AND NVENC's
+    /// RGB→YUV off the 3D engine so capture+encode don't fight a GPU-saturating game. Returns the YUV
+    /// texture, or `None` to fall back to the legacy RGB path (processor disabled/unavailable). Lazily
+    /// builds + caches the processor + output texture for the current size + HDR-ness.
+    unsafe fn convert_to_yuv(
+        &mut self,
+        input: &ID3D11Texture2D,
+        hdr: bool,
+    ) -> Option<ID3D11Texture2D> {
+        if self.vp_disabled {
+            return None;
+        }
+        if self.video_conv.is_none() || self.yuv_out.is_none() || self.yuv_is_hdr != hdr {
+            self.video_conv = None;
+            self.yuv_out = None;
+            let vc = match VideoConverter::new(
+                &self.device,
+                &self.context,
+                self.width,
+                self.height,
+                hdr,
+            ) {
+                Ok(vc) => vc,
+                Err(e) => {
+                    tracing::warn!(error = %format!("{e:#}"),
+                            "video processor unavailable — falling back to RGB encode path");
+                    self.vp_disabled = true;
+                    return None;
+                }
+            };
+            let fmt = if hdr {
+                windows::Win32::Graphics::Dxgi::Common::DXGI_FORMAT_P010
+            } else {
+                windows::Win32::Graphics::Dxgi::Common::DXGI_FORMAT_NV12
+            };
+            let desc = D3D11_TEXTURE2D_DESC {
+                Width: self.width,
+                Height: self.height,
+                MipLevels: 1,
+                ArraySize: 1,
+                Format: fmt,
+                SampleDesc: DXGI_SAMPLE_DESC {
+                    Count: 1,
+                    Quality: 0,
+                },
+                Usage: D3D11_USAGE_DEFAULT,
+                BindFlags: D3D11_BIND_RENDER_TARGET.0 as u32,
+                CPUAccessFlags: 0,
+                MiscFlags: 0,
+            };
+            let mut t: Option<ID3D11Texture2D> = None;
+            if let Err(e) = self.device.CreateTexture2D(&desc, None, Some(&mut t)) {
+                tracing::warn!(error = %format!("{e:?}"),
+                    "CreateTexture2D(YUV out) failed — falling back to RGB encode path");
+                self.vp_disabled = true;
+                return None;
+            }
+            self.video_conv = Some(vc);
+            self.yuv_out = t;
+            self.yuv_is_hdr = hdr;
+            tracing::info!(
+                hdr,
+                "video-processor YUV path active ({} on the video engine, 0% 3D)",
+                if hdr { "P010" } else { "NV12" }
+            );
+        }
+        let out = self.yuv_out.clone()?;
+        if let Err(e) = self.video_conv.as_ref()?.convert(input, &out) {
+            tracing::warn!(error = %format!("{e:#}"),
+                "VideoProcessorBlt failed — falling back to RGB encode path");
+            self.vp_disabled = true;
+            self.video_conv = None;
+            self.yuv_out = None;
+            return None;
+        }
+        Some(out)
+    }
+
    /// FP16 (`R16G16B16A16_FLOAT`) copy of the HDR duplication surface (RT for the cursor composite +
    /// SRV for the converter). Reallocated when absent (device/size change drops it).
    unsafe fn ensure_fp16_src(&mut self) -> Result<()> {
@@ -1718,6 +2050,9 @@ impl DuplCapturer {
        self.fp16_srv = None;
        self.hdr10_out = None;
        self.hdr_conv = None;
+        // Video processor + its YUV output belonged to the old device / size / HDR-ness — rebuild lazily.
+        self.video_conv = None;
+        self.yuv_out = None;
        self.first_frame = true;
        // Capture the CURRENT desktop frame as `last_present` (instead of seeding black). The secure
        // (lock/login/UAC) desktop is STATIC, so DDA only emits a frame on change — if we seeded black
@@ -1982,6 +2317,22 @@ impl DuplCapturer {
            let _ = self.dupl.as_ref().map(|d| d.ReleaseFrame());
            self.holding_frame = false;
            self.composite_cursor_gpu(&src, true)?; // onto the FP16 surface (HDR: decode + nits scale)
+                                                    // Video-engine path: scRGB FP16 → BT.2020 PQ P010 on the VIDEO engine (no 3D shader, and
+                                                    // NVENC encodes P010 natively). Fall back to the HdrConverter pixel shader (3D) only if the
+                                                    // video processor is unavailable.
+            if let Some(p010) = self.convert_to_yuv(&src, true) {
+                self.last_present = Some((p010.clone(), PixelFormat::P010));
+                return Ok(CapturedFrame {
+                    width: self.width,
+                    height: self.height,
+                    pts_ns: now_ns(),
+                    format: PixelFormat::P010,
+                    payload: FramePayload::D3d11(D3d11Frame {
+                        texture: p010,
+                        device: self.device.clone(),
+                    }),
+                });
+            }
            self.ensure_hdr10_out()?;
            let out = self.hdr10_out.clone().context("hdr10 out texture")?;
            if self.hdr_conv.is_none() {
@@ -2014,12 +2365,34 @@ impl DuplCapturer {
        if self.gpu_mode {
            // Zero-copy path: keep the frame on the GPU for NVENC. Copy the transient duplication
            // surface into a reused owned texture, release the duplication frame, hand off the texture.
+            // NOTE: do NOT convert the duplication surface directly on the video processor to skip this
+            // copy — the VP colour-convert (3D/compute on NVIDIA) holds the DDA surface until it
+            // completes, blocking ReleaseFrame/AcquireNextFrame and SERIALIZING capture+convert (~60 fps,
+            // encode_us 15-20 ms measured). The fast same-format CopyResource decouples them: it releases
+            // the DDA frame immediately so the convert runs independently (40-200 fps). Worth ~5% 3D.
            self.ensure_gpu_copy()?;
            let gpu = self.gpu_copy.clone().context("gpu copy texture")?;
            self.context.CopyResource(&gpu, &tex);
            let _ = self.dupl.as_ref().map(|d| d.ReleaseFrame());
            self.holding_frame = false;
            self.composite_cursor_gpu(&gpu, false)?;
+            // Prefer the video-engine YUV path (BGRA → NV12 on the video engine) so the colour
+            // conversion AND NVENC's encode stay OFF the 3D engine — the only way to keep up when a
+            // game pins the 3D engine at ~100%. Fall back to handing NVENC the BGRA texture (it then
+            // does RGB→YUV internally on the 3D/compute engine).
+            if let Some(nv12) = self.convert_to_yuv(&gpu, false) {
+                self.last_present = Some((nv12.clone(), PixelFormat::Nv12));
+                return Ok(CapturedFrame {
+                    width: self.width,
+                    height: self.height,
+                    pts_ns: now_ns(),
+                    format: PixelFormat::Nv12,
+                    payload: FramePayload::D3d11(D3d11Frame {
+                        texture: nv12,
+                        device: self.device.clone(),
+                    }),
+                });
+            }
            self.last_present = Some((gpu.clone(), PixelFormat::Bgra));
            return Ok(CapturedFrame {
                width: self.width,
@@ -17,10 +17,12 @@
 //! the DDA backend ([`super::dxgi::DuplCapturer`]) for those (see capture.rs).

 use super::dxgi::{
-    find_output, make_device, nudge_cursor_onto, D3d11Frame, HdrConverter, WinCaptureTarget,
+    find_output, make_device, nudge_cursor_onto, D3d11Frame, HdrConverter, VideoConverter,
+    WinCaptureTarget,
 };
 use super::{CapturedFrame, Capturer, FramePayload, PixelFormat};
 use anyhow::{bail, Context, Result};
+use std::collections::VecDeque;
 use std::sync::atomic::{AtomicU64, Ordering};
 use std::sync::{Arc, Condvar, Mutex};
 use std::time::{Duration, Instant};
@@ -37,8 +39,8 @@ use windows::Win32::Graphics::Direct3D11::{
    D3D11_USAGE_DEFAULT,
 };
 use windows::Win32::Graphics::Dxgi::Common::{
-    DXGI_COLOR_SPACE_RGB_FULL_G2084_NONE_P2020, DXGI_FORMAT_B8G8R8A8_UNORM,
-    DXGI_FORMAT_R10G10B10A2_UNORM, DXGI_FORMAT_R16G16B16A16_FLOAT, DXGI_SAMPLE_DESC,
+    DXGI_COLOR_SPACE_RGB_FULL_G2084_NONE_P2020, DXGI_FORMAT_R10G10B10A2_UNORM,
+    DXGI_FORMAT_R16G16B16A16_FLOAT, DXGI_SAMPLE_DESC,
 };
 use windows::Win32::Graphics::Dxgi::{IDXGIDevice, IDXGIOutput6};
 use windows::Win32::Security::{ImpersonateLoggedOnUser, RevertToSelf};
@@ -49,6 +51,22 @@ use windows::Win32::System::WinRT::Direct3D11::{
 use windows::Win32::System::WinRT::Graphics::Capture::IGraphicsCaptureItemInterop;
 use windows::Win32::System::WinRT::{RoInitialize, RO_INIT_MULTITHREADED};

+/// Output texture ring depth. The encode loop pipelines one frame deep (NVENC encodes frame N while
+/// the capturer produces N+1), so two live textures suffice; three gives headroom against a slow
+/// `lock_bitstream` and matches the WGC frame-pool depth.
+// Sized for the deep encode pipeline (`PUNKTFUNK_ENCODE_DEPTH`, default 4, clamped ≤ 6): up to DEPTH
+// frames are in flight in NVENC at once, so the HDR convert ring and the SDR held-frame set must each
+// keep DEPTH(+headroom) live textures, and the WGC pool needs spare buffers beyond what we hold.
+const OUT_RING: usize = 8;
+
+/// SDR zero-copy: how many recent WGC frames to keep alive so NVENC can encode the pool texture in
+/// place (no `CopyResource`). Each in-flight encode reads a distinct frame, so this must exceed the
+/// pipeline depth; the oldest is released once `HELD_FRAMES` newer ones exist.
+const HELD_FRAMES: usize = 8;
+/// WGC frame-pool buffer count. Must exceed `HELD_FRAMES` so the compositor always has free buffers
+/// to render into while we hold frames for in-place (zero-copy) SDR encode.
+const WGC_POOL_BUFFERS: i32 = 10;
+
 /// The host runs as SYSTEM (so the DDA secure-desktop path works), but WGC will NOT activate under
 /// the SYSTEM account (`CreateForMonitor` → 0x80070424). Impersonate the interactive console user
 /// for the WGC activation. Returns the user token (the caller reverts + closes it after activation)
@@ -112,8 +130,27 @@ pub struct WgcCapturer {
    hdr_conv: Option<HdrConverter>,
    fp16_src: Option<ID3D11Texture2D>,
    fp16_srv: Option<ID3D11ShaderResourceView>,
-    hdr10_out: Option<ID3D11Texture2D>,
-    bgra_copy: Option<ID3D11Texture2D>,
+    /// Ring of host-owned output textures (BGRA for SDR, R10G10B10A2 for HDR), rotated per processed
+    /// frame. A ring — not one texture — is required because the encode loop is PIPELINED: NVENC
+    /// encodes frame N (in place, registered by pointer) while this capturer produces frame N+1, so
+    /// N+1 must land in a DIFFERENT texture or it clobbers the in-flight encode. (`fp16_src` stays
+    /// single: it's only touched within the D3D11 immediate context, whose op ordering already
+    /// serializes the convert's read against the next copy's write — NVENC's async engine read is the
+    /// only consumer that escapes that ordering, and it reads the ring output, never `fp16_src`.)
+    out_ring: Vec<ID3D11Texture2D>,
+    ring_idx: usize,
+    /// Video-processor RGB→YUV converter (off the 3D engine where possible) + its NV12/P010 output
+    /// ring. Preferred path: the OS-composited capture (cursor already in it) is converted DIRECTLY to
+    /// NVENC's native YUV — no `CopyResource`, no cursor draw, and NVENC skips its internal RGB→YUV.
+    /// `None`/error → falls back to the legacy SDR-zero-copy / HDR-shader paths.
+    video_conv: Option<VideoConverter>,
+    yuv_out: Vec<ID3D11Texture2D>,
+    yuv_idx: usize,
+    yuv_is_hdr: bool,
+    vp_disabled: bool,
+    /// SDR zero-copy: the recent WGC frames we hand to NVENC in place. Held so the pool doesn't
+    /// recycle the texture mid-encode; the oldest is released once `HELD_FRAMES` newer ones exist.
+    held: VecDeque<Direct3D11CaptureFrame>,
    /// Last presentable GPU texture + format, repeated when no new frame arrived (static desktop).
    last_present: Option<(ID3D11Texture2D, PixelFormat)>,

@@ -204,10 +241,15 @@ impl WgcCapturer {
            } else {
                DirectXPixelFormat::B8G8R8A8UIntNormalized
            };
-            // ≥3 buffers for 240 Hz headroom (avoid the producer waiting on a free buffer).
-            let pool =
-                Direct3D11CaptureFramePool::CreateFreeThreaded(&d3d_device, pixel_format, 3, size)
-                    .context("CreateFreeThreaded frame pool")?;
+            // Extra buffers: SDR zero-copy holds the last `HELD_FRAMES` frames (encoded in place), so
+            // the pool needs headroom beyond that for the producer to keep rendering at 240 Hz.
+            let pool = Direct3D11CaptureFramePool::CreateFreeThreaded(
+                &d3d_device,
+                pixel_format,
+                WGC_POOL_BUFFERS,
+                size,
+            )
+            .context("CreateFreeThreaded frame pool")?;

            let signal = Arc::new(WgcSignal {
                available: AtomicU64::new(0),
@@ -278,8 +320,14 @@ impl WgcCapturer {
                hdr_conv: None,
                fp16_src: None,
                fp16_srv: None,
-                hdr10_out: None,
-                bgra_copy: None,
+                out_ring: Vec::new(),
+                ring_idx: 0,
+                video_conv: None,
+                yuv_out: Vec::new(),
+                yuv_idx: 0,
+                yuv_is_hdr: false,
+                vp_disabled: std::env::var_os("PUNKTFUNK_NO_VIDEO_PROCESSOR").is_some(),
+                held: VecDeque::new(),
                last_present: None,
                _keepalive: None,
            })
@@ -347,38 +395,112 @@ impl WgcCapturer {
        Ok(())
    }

-    unsafe fn ensure_hdr10_out(&mut self) -> Result<()> {
-        if self.hdr10_out.is_none() {
-            let desc = tex_desc(
-                self.width,
-                self.height,
-                DXGI_FORMAT_R10G10B10A2_UNORM,
-                D3D11_BIND_RENDER_TARGET.0 as u32,
-            );
+    /// Lazily allocate the HDR output texture ring (R10G10B10A2, the convert pass's render target →
+    /// NVENC input), `RENDER_TARGET`-bindable. SDR is zero-copy (encodes the WGC pool texture in
+    /// place) and uses no ring.
+    unsafe fn ensure_out_ring(
+        &mut self,
+        format: windows::Win32::Graphics::Dxgi::Common::DXGI_FORMAT,
+    ) -> Result<()> {
+        if !self.out_ring.is_empty() {
+            return Ok(());
+        }
+        let desc = tex_desc(
+            self.width,
+            self.height,
+            format,
+            D3D11_BIND_RENDER_TARGET.0 as u32,
+        );
+        for _ in 0..OUT_RING {
            let mut t = None;
            self.device
                .CreateTexture2D(&desc, None, Some(&mut t))
-                .context("CreateTexture2D(wgc hdr10 out)")?;
-            self.hdr10_out = t;
+                .context("CreateTexture2D(wgc out ring)")?;
+            self.out_ring.push(t.context("wgc out ring tex")?);
        }
        Ok(())
    }

-    unsafe fn ensure_bgra(&mut self) -> Result<()> {
-        if self.bgra_copy.is_none() {
+    /// Convert `input` (the OS-composited WGC pool texture: BGRA or scRGB FP16) → NVENC's native YUV
+    /// (NV12 / P010) on the video processor. Returns the YUV texture (from a ring so consecutive
+    /// encodes don't collide), or `None` to fall back to the legacy RGB paths.
+    unsafe fn convert_to_yuv(
+        &mut self,
+        input: &ID3D11Texture2D,
+        hdr: bool,
+    ) -> Option<ID3D11Texture2D> {
+        if self.vp_disabled {
+            return None;
+        }
+        if self.video_conv.is_none() || self.yuv_out.is_empty() || self.yuv_is_hdr != hdr {
+            self.video_conv = None;
+            self.yuv_out.clear();
+            self.yuv_idx = 0;
+            let vc = match VideoConverter::new(
+                &self.device,
+                &self.context,
+                self.width,
+                self.height,
+                hdr,
+            ) {
+                Ok(vc) => vc,
+                Err(e) => {
+                    tracing::warn!(error = %format!("{e:#}"),
+                        "WGC: video processor unavailable — falling back to RGB path");
+                    self.vp_disabled = true;
+                    return None;
+                }
+            };
+            let fmt = if hdr {
+                windows::Win32::Graphics::Dxgi::Common::DXGI_FORMAT_P010
+            } else {
+                windows::Win32::Graphics::Dxgi::Common::DXGI_FORMAT_NV12
+            };
            let desc = tex_desc(
                self.width,
                self.height,
-                DXGI_FORMAT_B8G8R8A8_UNORM,
+                fmt,
                D3D11_BIND_RENDER_TARGET.0 as u32,
            );
-            let mut t = None;
-            self.device
-                .CreateTexture2D(&desc, None, Some(&mut t))
-                .context("CreateTexture2D(wgc bgra)")?;
-            self.bgra_copy = t;
+            for _ in 0..OUT_RING {
+                let mut t = None;
+                if self
+                    .device
+                    .CreateTexture2D(&desc, None, Some(&mut t))
+                    .is_err()
+                {
+                    tracing::warn!("WGC: CreateTexture2D(YUV) failed — falling back to RGB path");
+                    self.vp_disabled = true;
+                    self.yuv_out.clear();
+                    return None;
+                }
+                let Some(tex) = t else {
+                    self.vp_disabled = true;
+                    self.yuv_out.clear();
+                    return None;
+                };
+                self.yuv_out.push(tex);
+            }
+            self.video_conv = Some(vc);
+            self.yuv_is_hdr = hdr;
+            tracing::info!(
+                hdr,
+                "WGC: video-processor YUV path active ({})",
+                if hdr { "P010" } else { "NV12" }
+            );
        }
-        Ok(())
+        let slot = self.yuv_idx;
+        self.yuv_idx = (self.yuv_idx + 1) % self.yuv_out.len();
+        let out = self.yuv_out[slot].clone();
+        if let Err(e) = self.video_conv.as_ref()?.convert(input, &out) {
+            tracing::warn!(error = %format!("{e:#}"),
+                "WGC: VideoProcessorBlt failed — falling back to RGB path");
+            self.vp_disabled = true;
+            self.video_conv = None;
+            self.yuv_out.clear();
+            return None;
+        }
+        Some(out)
    }

    fn process_frame(&mut self, frame: Direct3D11CaptureFrame) -> Result<CapturedFrame> {
@@ -391,13 +513,38 @@ impl WgcCapturer {
                .GetInterface()
                .context("GetInterface ID3D11Texture2D")?;

+            // Preferred path: convert the OS-composited capture (cursor already in it) DIRECTLY to
+            // NVENC's native YUV on the video processor — no CopyResource, no cursor draw, and NVENC
+            // skips its internal RGB→YUV (the contended 3D step). WGC's multi-buffer pool + held set
+            // means reading the pool texture directly does NOT serialize (unlike DDA's single-frame
+            // model). The frame is held until the async Blt finishes.
+            if let Some(yuv) = self.convert_to_yuv(&src, self.hdr) {
+                let fmt = if self.hdr {
+                    PixelFormat::P010
+                } else {
+                    PixelFormat::Nv12
+                };
+                self.last_present = Some((yuv.clone(), fmt));
+                let out = self.d3d11_frame(yuv, fmt);
+                self.held.push_back(frame);
+                while self.held.len() > HELD_FRAMES {
+                    self.held.pop_front();
+                }
+                return Ok(out);
+            }
+
+            // --- fallback (video processor unavailable) ---
            if self.hdr {
+                // Next ring slot — the in-flight encode reads the slot we handed out last time, so
+                // this capture must land in a different one (see `out_ring`).
+                let slot = self.ring_idx;
+                self.ring_idx = (self.ring_idx + 1) % OUT_RING;
                // FP16 (cursor already composited by the OS) → BT.2020 PQ 10-bit for NVENC.
                self.ensure_fp16_src()?;
                let fp16 = self.fp16_src.clone().context("fp16 src")?;
                self.context.CopyResource(&fp16, &src);
-                self.ensure_hdr10_out()?;
-                let out = self.hdr10_out.clone().context("hdr10 out")?;
+                self.ensure_out_ring(DXGI_FORMAT_R10G10B10A2_UNORM)?;
+                let out = self.out_ring[slot].clone();
                if self.hdr_conv.is_none() {
                    self.hdr_conv = Some(HdrConverter::new(&self.device)?);
                }
@@ -416,12 +563,19 @@ impl WgcCapturer {
                self.last_present = Some((out.clone(), PixelFormat::Rgb10a2));
                Ok(self.d3d11_frame(out, PixelFormat::Rgb10a2))
            } else {
-                // SDR: copy out of the recycled pool texture (cursor already composited) and hand off.
-                self.ensure_bgra()?;
-                let bgra = self.bgra_copy.clone().context("bgra copy")?;
-                self.context.CopyResource(&bgra, &src);
-                self.last_present = Some((bgra.clone(), PixelFormat::Bgra));
-                Ok(self.d3d11_frame(bgra, PixelFormat::Bgra))
+                // SDR ZERO-COPY: hand NVENC the WGC pool texture DIRECTLY — no `CopyResource`. The
+                // per-frame copy otherwise queues on the graphics engine behind a GPU-saturating game
+                // and stalls `lock_bitstream` ~20 ms (NVENC sits idle waiting for its input). Encoding
+                // the pool texture in place removes that graphics-queue dependency (Apollo's model).
+                // We must keep the frame alive until its async encode finishes, so retain the last
+                // `HELD_FRAMES`; the pool has spare buffers so the producer never starves.
+                self.last_present = Some((src.clone(), PixelFormat::Bgra));
+                let out = self.d3d11_frame(src, PixelFormat::Bgra);
+                self.held.push_back(frame);
+                while self.held.len() > HELD_FRAMES {
+                    self.held.pop_front();
+                }
+                Ok(out)
            }
        }
    }
@@ -103,9 +103,10 @@ fn nvenc_input(format: PixelFormat) -> (Pixel, bool) {
        PixelFormat::Rgba => (Pixel::RGBA, false),
        PixelFormat::Rgb => (Pixel::RGBZ, true), // RGB -> rgb0
        PixelFormat::Bgr => (Pixel::BGRZ, true), // BGR -> bgr0
-        // 10-bit HDR (R10G10B10A2) is produced only by the Windows DXGI HDR capture path; the Linux
-        // capturer never emits it. Map to BGRA so the match is exhaustive — unreachable here.
-        PixelFormat::Rgb10a2 => (Pixel::BGRA, false),
+        // Rgb10a2 (HDR) and NV12/P010 (the Windows video-processor YUV outputs) are produced only by
+        // the Windows capture/encode paths; the Linux capturer never emits them. Map to BGRA so the
+        // match is exhaustive — unreachable here.
+        PixelFormat::Rgb10a2 | PixelFormat::Nv12 | PixelFormat::P010 => (Pixel::BGRA, false),
    }
 }

@@ -25,7 +25,10 @@ use windows::Win32::Graphics::Direct3D11::{ID3D11Device, ID3D11Texture2D};
 use nvidia_video_codec_sdk::sys::nvEncodeAPI as nv;
 use nvidia_video_codec_sdk::ENCODE_API as API;

-const POOL: usize = 4;
+// Output bitstream buffers = max in-flight encodes. The helper deep-pipelines (submits several frames
+// before locking the oldest) so per-frame GPU-scheduling waits OVERLAP instead of serializing under a
+// GPU-saturating game; this must be ≥ the helper's `PUNKTFUNK_ENCODE_DEPTH` (default 4, clamped ≤ 6).
+const POOL: usize = 8;

 fn codec_guid(codec: Codec) -> nv::GUID {
    match codec {
@@ -363,7 +366,9 @@ impl Encoder for NvencD3d11Encoder {
        // frame arrives on a different device OR at a different size than our session was built on.
        // HDR (BT.2020 PQ 10-bit) when the capturer hands us a 10-bit R10G10B10A2 frame. This can flip
        // mid-session when the user toggles HDR (which arrives as a capture device recreate anyway).
-        let hdr = matches!(captured.format, PixelFormat::Rgb10a2);
+        // HDR (BT.2020 PQ) when the capturer hands a 10-bit frame — either R10G10B10A2 (the legacy
+        // shader path) or P010 (the video-processor path). 8-bit NV12/ARGB → SDR.
+        let hdr = matches!(captured.format, PixelFormat::Rgb10a2 | PixelFormat::P010);
        let dev_raw = frame.device.as_raw();
        let size_changed =
            self.inited && (self.width != captured.width || self.height != captured.height);
@@ -384,13 +389,22 @@ impl Encoder for NvencD3d11Encoder {
            self.width = captured.width;
            self.height = captured.height;
            self.hdr = hdr;
-            if hdr {
-                // 10-bit BT.2020 PQ input; force Main10 regardless of the negotiated SDR bit depth.
-                self.bit_depth = 10;
-                self.buffer_fmt = nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ABGR10;
-            } else {
-                self.buffer_fmt = nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ARGB;
-            }
+            // Pick the NVENC input format from the captured pixel format. YUV (NV12/P010) is the
+            // video-processor path — NVENC encodes it natively (no internal RGB→YUV, which is a hidden
+            // 3D/compute step that would fight a GPU-saturating game). RGB (ARGB/ABGR10) is the legacy
+            // shader path. 10-bit (P010/ABGR10) forces HEVC Main10 + the BT.2020 PQ VUI.
+            self.buffer_fmt = match captured.format {
+                PixelFormat::P010 => {
+                    self.bit_depth = 10;
+                    nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_YUV420_10BIT
+                }
+                PixelFormat::Rgb10a2 => {
+                    self.bit_depth = 10;
+                    nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ABGR10
+                }
+                PixelFormat::Nv12 => nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_NV12,
+                _ => nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ARGB,
+            };
            let device = frame.device.clone();
            self.init_session(&device)?;
            self.init_device = dev_raw;
@@ -151,6 +151,13 @@ impl Encoder for OpenH264Encoder {
            PixelFormat::Rgb10a2 => {
                anyhow::bail!("software H.264 encoder cannot encode 10-bit HDR (Rgb10a2)")
            }
+            // NV12/P010 are GPU-resident video-processor outputs for the NVENC path; the software
+            // encoder never receives them (it only gets CPU RGB frames).
+            PixelFormat::Nv12 | PixelFormat::P010 => {
+                anyhow::bail!(
+                    "software encoder cannot encode YUV GPU textures (NV12/P010 → NVENC only)"
+                )
+            }
        }

        if self.force_kf {
@@ -101,19 +101,85 @@ pub fn run(opts: HelperOptions) -> Result<()> {
    let stdout = std::io::stdout();
    let mut out = stdout.lock();

-    let mut frame = first;
+    // Encode pipeline depth. The loop keeps DEPTH frames in flight so per-frame GPU-scheduling waits
+    // can overlap. NOTE: depth > 1 was measured to REGRESS under a GPU-saturating game — the encodes
+    // serialize on the contended GPU anyway, so a deeper queue just stacks latency (≈ depth × frame
+    // time) without raising throughput. Default 1 (the validated-best); `PUNKTFUNK_ENCODE_DEPTH` (1..=6)
+    // can raise it if a future workload is genuinely encode-throughput-bound rather than scheduling-bound.
+    let depth: usize = std::env::var("PUNKTFUNK_ENCODE_DEPTH")
+        .ok()
+        .and_then(|s| s.trim().parse::<usize>().ok())
+        .filter(|&d| (1..=6).contains(&d))
+        .unwrap_or(1);
+    tracing::info!(depth, "WGC helper: encode pipeline depth");
+
+    let perf = std::env::var_os("PUNKTFUNK_PERF").is_some();
+    let mut frames = 0u64;
+    let mut cap_wait_ns = 0u64;
+    let mut encode_ns = 0u64; // time blocked in lock_bitstream (the oldest in-flight encode)
+    let mut write_ns = 0u64; // time blocked writing the AU to the stdout pipe (relay backpressure)
+    let mut window = std::time::Instant::now();
+
+    // Prime: submit `depth` frames before the first poll so NVENC has that many encodes in flight.
+    // We don't hold the `CapturedFrame`s past `submit`: NVENC keeps its own registered texture clone
+    // and the capturer's ring/held-set own the canonical refs (sized for `depth`), so the in-flight
+    // inputs stay valid after our clones drop.
+    enc.submit(&first).context("first encoder submit")?;
+    drop(first);
+    for _ in 1..depth {
+        let f = cap.next_frame().context("WGC prime frame")?;
+        enc.submit(&f).context("prime encoder submit")?;
+    }
    loop {
        if kf.swap(false, Ordering::Relaxed) {
            enc.request_keyframe();
        }
-        enc.submit(&frame).context("encoder submit")?;
-        while let Some(au) = enc.poll().context("encoder poll")? {
-            if write_au(&mut out, &au).is_err() {
+        // Pop + forward the OLDEST in-flight frame (FIFO). With `depth` outstanding it has had
+        // depth-1 frames' worth of GPU slots to finish, so this rarely blocks under load.
+        let p0 = std::time::Instant::now();
+        let polled = enc.poll().context("encoder poll")?;
+        if perf {
+            encode_ns += p0.elapsed().as_nanos() as u64;
+        }
+        if let Some(au) = polled {
+            let w0 = std::time::Instant::now();
+            let wrote = write_au(&mut out, &au);
+            if perf {
+                write_ns += w0.elapsed().as_nanos() as u64;
+            }
+            if wrote.is_err() {
                tracing::info!("WGC helper: stdout closed (host gone) — exiting");
                return Ok(());
            }
        }
-        frame = cap.next_frame().context("WGC next frame")?;
+        // Refill: capture + submit to keep `depth` frames in flight.
+        let t0 = std::time::Instant::now();
+        let next = cap.next_frame().context("WGC next frame")?;
+        if perf {
+            cap_wait_ns += t0.elapsed().as_nanos() as u64;
+        }
+        enc.submit(&next).context("encoder submit")?;
+
+        if perf {
+            frames += 1;
+            let since = window.elapsed();
+            if since.as_secs() >= 2 {
+                let secs = since.as_secs_f64();
+                let per = |ns: u64| format!("{:.2}", ns as f64 / frames as f64 / 1e6);
+                tracing::info!(
+                    fps = format!("{:.1}", frames as f64 / secs),
+                    cap_wait_ms = per(cap_wait_ns),
+                    encode_ms = per(encode_ns),
+                    write_ms = per(write_ns),
+                    "WGC helper perf (depth-pipelined; encode_ms=lock_bitstream on the oldest)"
+                );
+                frames = 0;
+                cap_wait_ns = 0;
+                encode_ns = 0;
+                write_ns = 0;
+                window = std::time::Instant::now();
+            }
+        }
    }
 }

@@ -36,8 +36,8 @@ pub fn drm_fourcc(format: crate::capture::PixelFormat) -> Option<u32> {
        Rgbx => fourcc(b"XB24"), // DRM_FORMAT_XBGR8888
        Rgba => fourcc(b"AB24"), // DRM_FORMAT_ABGR8888
        // 24-bit packed RGB/BGR have no straightforward dmabuf import here; use the CPU path.
-        // Rgb10a2 is the Windows HDR capture format — never produced by the Linux capturer.
-        Rgb | Bgr | Rgb10a2 => return None,
+        // Rgb10a2/Nv12/P010 are the Windows HDR / video-processor formats — never produced on Linux.
+        Rgb | Bgr | Rgb10a2 | Nv12 | P010 => return None,
    })
 }