fix(pf-encode): arbitrate NVENC split-encode vs sub-frame readback (Phase 8)

Verified against nvEncodeAPI.h's own splitEncodeMode doc (user-prompted
— the audit's 'exclusive for HEVC' one-liner deserved checking):
- H.264: split 'is not applicable' — hard-DISABLE the mode so the
  written config, CeilingKey, the split diagnostic log and the
  rejection-retry stay truthful (the retry used to re-open a
  byte-identical session after an H.264 'split rejection'). The libav
  path's operator arm gains the codec gate its auto arm always had.
- HEVC: split 'not supported if … subframe mode' — when WE force split
  (TWO/THREE/AUTO_FORCED, the 4K120 throughput lever), sub-frame yields
  with a logged escape (PUNKTFUNK_SPLIT_ENCODE=0 chooses sub-frame).
  ⚠ Keyed on FORCED modes only, never != DISABLE: AUTO(0) is the
  resolver's fallthrough for every sub-950Mpix session, and the wider
  key would have disarmed the Phase-3 chunked-poll feature fleet-wide
  (critic catch). Under AUTO the driver arbitrates — the shipped state.
- AV1: untouched — per-tile sub-frame + split are legal together.

The arbitration is a pure nvenc_core fn called by each backend BEFORE
the ladder, the ceiling key and the chunked-poll latch — all three see
the post-arbitration truth. A drop inside build_init_params would have
left poll_chunk busy-polling its whole budget every AU (numSlices stays
0 without reportSliceOffsets; both loop exits dead — critic catch).
Linux latches subframe_forced beside subframe_on at query_caps (no env
re-reads after open); Windows records the arbitrated state so
reconfigure presents exactly the params the open had (also closes the
pre-existing mid-session env-flip hazard there).

Truth-table tests in nvenc_core; PUNKTFUNK_NVENC_SUBFRAME documented
(it never was); PUNKTFUNK_SPLIT_ENCODE row updated.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

crates/pf-encode/src/enc/linux/mod.rs

@@ -448,7 +448,17 @@ impl NvencEncoder {
         let pix_rate = width as u64 * height as u64 * fps as u64;
         let split = std::env::var("PUNKTFUNK_SPLIT_ENCODE").ok();
         match split.as_deref() {
-            Some(mode) => opts.set("split_encode_mode", mode),
+            // The operator arm gains the codec gate the auto arm always had (Phase 8): split
+            // "is not applicable to H264" per nvEncodeAPI.h, and h264_nvenc has no such AVOption
+            // — setting it would fail the open on a leftover dict entry.
+            Some(mode) if matches!(codec, Codec::H265 | Codec::Av1) => {
+                opts.set("split_encode_mode", mode)
+            }
+            Some(_) => tracing::warn!(
+                codec = codec.nvenc_name(),
+                "PUNKTFUNK_SPLIT_ENCODE ignored — split encoding is not applicable to H.264 \
+                 (nvEncodeAPI.h)"
+            ),
             None if matches!(codec, Codec::H265 | Codec::Av1)
                 && pix_rate >= super::SPLIT_FORCE_PIXEL_RATE =>
             {

crates/pf-encode/src/enc/linux/nvenc_cuda.rs

@@ -62,8 +62,8 @@
 
 use super::nvenc_core::{
     apply_low_latency_config, build_init_params, cached_ceiling, codec_guid, plan_range_recovery,
-    resolve_slices, resolve_split_mode, resolve_subframe, store_ceiling, CeilingKey,
-    LowLatencyConfig, NvStatusExt, RangePlan,
+    resolve_slices, resolve_split_mode, resolve_split_subframe, resolve_subframe, store_ceiling,
+    subframe_env_forced, CeilingKey, LowLatencyConfig, NvStatusExt, RangePlan,
 };
 use super::nvenc_status;
 use super::{AuChunk, ChromaFormat, Codec, EncodedFrame, Encoder, EncoderCaps};
@@ -629,6 +629,10 @@ pub struct NvencCudaEncoder {
     /// [`subframe_cap`](Self::subframe_cap)); consumed by every `build_init_params` call so the
     /// open and the in-place reconfigure present identical init params.
     subframe_on: bool,
+    /// Whether `PUNKTFUNK_NVENC_SUBFRAME=1` was EXPLICITLY forced — latched with `subframe_on`
+    /// (same invariant: open and reconfigure must present identical init params, so no env
+    /// re-reads after `query_caps`). Only consumed by [`resolve_split_subframe`]'s log severity.
+    subframe_forced: bool,
     /// Sub-frame chunked poll armed for the live session (§7 LN1 Phase 1): multi-slice +
     /// sub-frame readback configured AND sync retrieve at init. See [`Encoder::poll_chunk`].
     subframe_chunks: bool,
@@ -721,6 +725,7 @@ impl NvencCudaEncoder {
             slices: 1,
             subframe_cap: false,
             subframe_on: false,
+            subframe_forced: false,
             subframe_chunks: false,
             chunk: None,
         })
@@ -921,6 +926,7 @@ impl NvencCudaEncoder {
         // escapes.
         self.slices = resolve_slices(self.codec, 4);
         self.subframe_on = resolve_subframe(self.subframe_cap);
+        self.subframe_forced = subframe_env_forced();
         tracing::info!(
             rfi = self.rfi_supported,
             custom_vbv = self.custom_vbv,
@@ -1136,6 +1142,16 @@ impl NvencCudaEncoder {
             // [`resolve_split_mode`] for the precedence (env override / 10-bit / pixel rate).
             let pixel_rate = self.width as u64 * self.height as u64 * self.fps.max(1) as u64;
             let split_mode: u32 = resolve_split_mode(self.bit_depth, pixel_rate);
+            // Split × sub-frame arbitration (Phase 8) BEFORE the ladder, the ceiling key and the
+            // chunked-poll latch — all three must see the post-arbitration truth (a drop inside
+            // build_init_params would leave poll_chunk busy-polling its whole budget per AU).
+            let (split_mode, subframe_on) = resolve_split_subframe(
+                self.codec,
+                split_mode,
+                self.subframe_on,
+                self.subframe_forced,
+            );
+            self.subframe_on = subframe_on;
             const CLAMP_TOL_BPS: u64 = 20_000_000;
 
             // Ceiling cache (process lifetime, `nvenc_core`): a prior clamp search already found

crates/pf-encode/src/enc/nvenc_core.rs

@@ -102,6 +102,141 @@ pub(super) fn resolve_split_mode(bit_depth: u8, pixel_rate: u64) -> u32 {
     mode
 }
 
+/// Whether the operator EXPLICITLY forced sub-frame readback on (`PUNKTFUNK_NVENC_SUBFRAME=1`)
+/// — the log-severity input to [`resolve_split_subframe`]: a forced knob being overridden
+/// deserves a `warn`, a default being tuned an `info`. Callers LATCH this once next to their
+/// resolved subframe state (an env re-read at reconfigure would violate the "open and
+/// reconfigure present identical init params" invariant).
+/// Linux-cfg'd like its ONLY caller (the `nvenc_cuda` query_caps latch) — Windows sessions have
+/// `subframe == forced` by construction (env opt-in only) and never consult this; without the
+/// cfg it is dead code on every Windows leg (item-level dead_code, the recurring trap).
+#[cfg(target_os = "linux")]
+pub(super) fn subframe_env_forced() -> bool {
+    matches!(
+        std::env::var("PUNKTFUNK_NVENC_SUBFRAME").as_deref(),
+        Ok("1")
+    )
+}
+
+/// The split-encode × sub-frame arbitration (Phase 8; verified against `nvEncodeAPI.h`'s own
+/// `splitEncodeMode` doc, not folklore):
+/// - **H.264**: split "is not applicable" — hard-DISABLE the mode so the written config, the
+///   ceiling-cache key, the split diagnostic log and the rejection-retry all stay truthful (the
+///   retry used to re-open a byte-identical session after an H.264 "split rejection").
+/// - **HEVC**: split is "not supported if … subframe mode" — when WE force split
+///   (TWO/THREE/AUTO_FORCED, e.g. the 4K120 throughput requirement), sub-frame yields. Under
+///   plain AUTO the driver arbitrates — the shipped fleet state (1080p–1440p240 all run
+///   AUTO+subframe); keying on `!= DISABLE` here would have disarmed the Phase-3 chunked-poll
+///   feature fleet-wide.
+/// - **AV1**: both legal (sub-frame is per-tile; split is constrained only by
+///   output-into-vidmem, which we never use) — untouched.
+///
+/// Returns the `(split_mode, subframe)` to ACTUALLY configure. The caller must store BOTH back
+/// (the chunked-poll latch and `CeilingKey` key on them) — a silent in-params drop would leave
+/// `poll_chunk` busy-polling its full budget every AU (`numSlices` stays 0 without
+/// `reportSliceOffsets`, so neither loop exit ever fires).
+pub(super) fn resolve_split_subframe(
+    codec: Codec,
+    split_mode: u32,
+    subframe: bool,
+    subframe_forced: bool,
+) -> (u32, bool) {
+    use nv::NV_ENC_SPLIT_ENCODE_MODE as M;
+    if codec == Codec::H264 {
+        return (M::NV_ENC_SPLIT_DISABLE_MODE as u32, subframe);
+    }
+    let split_forced = split_mode == M::NV_ENC_SPLIT_TWO_FORCED_MODE as u32
+        || split_mode == M::NV_ENC_SPLIT_THREE_FORCED_MODE as u32
+        || split_mode == M::NV_ENC_SPLIT_AUTO_FORCED_MODE as u32;
+    if codec == Codec::H265 && split_forced && subframe {
+        if subframe_forced {
+            tracing::warn!(
+                split_mode,
+                "HEVC forced split-encode and PUNKTFUNK_NVENC_SUBFRAME=1 are mutually \
+                 unsupported (nvEncodeAPI.h) — sub-frame readback disabled for this session; \
+                 set PUNKTFUNK_SPLIT_ENCODE=0 to choose sub-frame instead"
+            );
+        } else {
+            tracing::info!(
+                split_mode,
+                "HEVC forced split-encode supersedes default-on sub-frame readback (mutually \
+                 unsupported per nvEncodeAPI.h; split is the 4K120 throughput lever) — set \
+                 PUNKTFUNK_SPLIT_ENCODE=0 to choose sub-frame instead"
+            );
+        }
+        return (split_mode, false);
+    }
+    (split_mode, subframe)
+}
+
+#[cfg(test)]
+mod split_subframe_tests {
+    use super::{resolve_split_subframe, Codec};
+    use nvidia_video_codec_sdk::sys::nvEncodeAPI::NV_ENC_SPLIT_ENCODE_MODE as M;
+
+    const AUTO: u32 = M::NV_ENC_SPLIT_AUTO_MODE as u32;
+    const TWO: u32 = M::NV_ENC_SPLIT_TWO_FORCED_MODE as u32;
+    const AUTO_F: u32 = M::NV_ENC_SPLIT_AUTO_FORCED_MODE as u32;
+    const DISABLE: u32 = M::NV_ENC_SPLIT_DISABLE_MODE as u32;
+
+    /// THE FLEET CASE: plain AUTO + default-on sub-frame must pass through untouched — the
+    /// driver arbitrates. Keying the rule on `!= DISABLE` would disarm sub-frame on every
+    /// default Linux HEVC session (AUTO == 0 is the resolver's fallthrough).
+    #[test]
+    fn hevc_auto_keeps_subframe() {
+        assert_eq!(
+            resolve_split_subframe(Codec::H265, AUTO, true, false),
+            (AUTO, true)
+        );
+    }
+
+    #[test]
+    fn hevc_forced_split_drops_subframe() {
+        assert_eq!(
+            resolve_split_subframe(Codec::H265, TWO, true, false),
+            (TWO, false)
+        );
+        assert_eq!(
+            resolve_split_subframe(Codec::H265, AUTO_F, true, true),
+            (AUTO_F, false)
+        );
+        // No sub-frame to drop → nothing changes.
+        assert_eq!(
+            resolve_split_subframe(Codec::H265, TWO, false, false),
+            (TWO, false)
+        );
+        // Explicitly disabled split → sub-frame kept (the documented escape).
+        assert_eq!(
+            resolve_split_subframe(Codec::H265, DISABLE, true, true),
+            (DISABLE, true)
+        );
+    }
+
+    /// H.264: split "is not applicable" (nvEncodeAPI.h) — hard-DISABLE regardless of the
+    /// resolved mode; sub-frame (H.264 slices) is unaffected.
+    #[test]
+    fn h264_split_hard_disabled() {
+        assert_eq!(
+            resolve_split_subframe(Codec::H264, TWO, true, false),
+            (DISABLE, true)
+        );
+        assert_eq!(
+            resolve_split_subframe(Codec::H264, AUTO, false, false),
+            (DISABLE, false)
+        );
+    }
+
+    /// AV1: both features are legal together (per-tile sub-frame; split constrained only by
+    /// output-into-vidmem) — the arbitration must not touch it.
+    #[test]
+    fn av1_untouched() {
+        assert_eq!(
+            resolve_split_subframe(Codec::Av1, TWO, true, true),
+            (TWO, true)
+        );
+    }
+}
+
 /// One session config's identity for the process-lifetime bitrate-ceiling cache
 /// ([`cached_ceiling`]/[`store_ceiling`]). Everything the driver's codec-level validation keys
 /// off: the GPU (different NVENC generations have different level ceilings), dims/fps (the luma

crates/pf-encode/src/enc/windows/nvenc.rs

@@ -38,8 +38,8 @@
 
 use super::nvenc_core::{
     apply_low_latency_config, build_init_params, cached_ceiling, codec_guid, plan_range_recovery,
-    resolve_slices, resolve_split_mode, resolve_subframe, store_ceiling, CeilingKey,
-    LowLatencyConfig, NvStatusExt, RangePlan,
+    resolve_slices, resolve_split_mode, resolve_split_subframe, resolve_subframe, store_ceiling,
+    CeilingKey, LowLatencyConfig, NvStatusExt, RangePlan,
 };
 use super::nvenc_status;
 use super::{ChromaFormat, Codec, EncodedFrame, Encoder, EncoderCaps};
@@ -603,6 +603,11 @@ pub struct NvencD3d11Encoder {
     /// `reconfigure_bitrate` must present the SAME init params as the open (only the config's
     /// rate fields may move). Meaningless while `inited` is false.
     split_mode: u32,
+    /// The session's arbitrated sub-frame state ([`resolve_split_subframe`] at open) — recorded
+    /// so reconfigure presents EXACTLY the init params the session opened with (an env re-read
+    /// there could flip `enableSubFrameWrite` mid-session, and would re-log the arbitration on
+    /// every ABR retarget).
+    subframe_on: bool,
     session_async: bool,
     /// The last reference-frame range we invalidated — dedupes repeated RFI requests for the same
     /// loss event (the client resends until it sees recovery).
@@ -686,6 +691,7 @@ impl NvencD3d11Encoder {
             rfi_supported: false,
             custom_vbv: false,
             split_mode: nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32,
+            subframe_on: false,
             session_async: false,
             last_rfi_range: None,
             init_device: ptr::null_mut(),
@@ -980,6 +986,7 @@ impl NvencD3d11Encoder {
         bitrate: u64,
         split_mode: u32,
         enable_async: bool,
+        subframe: bool,
     ) -> Result<*mut c_void> {
         let mut params = nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS {
             version: nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER,
@@ -1014,9 +1021,10 @@ impl NvencD3d11Encoder {
             &mut cfg,
             split_mode,
             enable_async,
-            // Windows: env opt-in only ("1"), never a default — and build_init_params
-            // additionally refuses it on an async session.
-            resolve_subframe(false),
+            // Windows: env opt-in only ("1"), never a default; arbitrated against split by the
+            // caller (resolve_split_subframe) — and build_init_params additionally refuses it
+            // on an async session.
+            subframe,
         );
 
         match (api().initialize_encoder)(enc, &mut init).nv_ok() {
@@ -1069,6 +1077,11 @@ impl NvencD3d11Encoder {
             // The init-failure fallback below disables it if a codec/config rejects it.
             let pixel_rate = self.width as u64 * self.height as u64 * self.fps.max(1) as u64;
             let split_mode: u32 = resolve_split_mode(self.bit_depth, pixel_rate);
+            // Split × sub-frame arbitration (Phase 8) before the ladder/ceiling key. On Windows
+            // sub-frame is env-opt-in only, so resolved == forced by construction.
+            let subframe_req = resolve_subframe(false);
+            let (split_mode, subframe_req) =
+                resolve_split_subframe(self.codec, split_mode, subframe_req, subframe_req);
             // Find the highest bitrate the GPU's codec LEVEL accepts and CLAMP to it. NVENC rejects
             // `initialize_encoder` (InvalidParam) when the bitrate exceeds the level ceiling (e.g. a
             // 1 Gbps request on HEVC). Strategy: try the requested rate; if the only problem is a forced
@@ -1098,12 +1111,19 @@ impl NvencD3d11Encoder {
                 }
             }
 
-            let mut probe = self.try_open_session(device, target_bps, split_mode, use_async);
+            let mut probe =
+                self.try_open_session(device, target_bps, split_mode, use_async, subframe_req);
             // The cache is advisory: a stale entry (driver change, identity collision) must not
             // wedge the open — retry the requested rate and let the search below rediscover.
             if probe.is_err() && target_bps < requested_bps {
                 target_bps = requested_bps;
-                probe = self.try_open_session(device, requested_bps, split_mode, use_async);
+                probe = self.try_open_session(
+                    device,
+                    requested_bps,
+                    split_mode,
+                    use_async,
+                    subframe_req,
+                );
             }
             // Disambiguate a forced-split rejection from a bitrate-cap rejection: retry once at the
             // requested rate with split disabled — if THAT succeeds, split was the problem, not bitrate.
@@ -1117,7 +1137,9 @@ impl NvencD3d11Encoder {
                 split_mode != nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
             if probe.is_err() && split_on {
                 let no_split = nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
-                if let Ok(e) = self.try_open_session(device, target_bps, no_split, use_async) {
+                if let Ok(e) =
+                    self.try_open_session(device, target_bps, no_split, use_async, subframe_req)
+                {
                     tracing::warn!("NVENC: split-encode rejected by codec/config — disabled");
                     used_split = no_split;
                     probe = Ok(e);
@@ -1144,7 +1166,13 @@ impl NvencD3d11Encoder {
                     let mut best_bps = 0u64;
                     while hi > lo + CLAMP_TOL_BPS {
                         let mid = lo + (hi - lo) / 2;
-                        match self.try_open_session(device, mid, used_split, use_async) {
+                        match self.try_open_session(
+                            device,
+                            mid,
+                            used_split,
+                            use_async,
+                            subframe_req,
+                        ) {
                             Ok(e) => {
                                 if !best.is_null() {
                                     let _ = (api().destroy_encoder)(best);
@@ -1169,12 +1197,23 @@ impl NvencD3d11Encoder {
                         // trying split-disabled in case a forced split (not the bitrate) is the blocker.
                         let no_split =
                             nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
-                        best = match self.try_open_session(device, FLOOR_BPS, used_split, use_async)
-                        {
+                        best = match self.try_open_session(
+                            device,
+                            FLOOR_BPS,
+                            used_split,
+                            use_async,
+                            subframe_req,
+                        ) {
                             Ok(e) => e,
                             Err(_) => {
                                 let e = self
-                                    .try_open_session(device, FLOOR_BPS, no_split, use_async)
+                                    .try_open_session(
+                                        device,
+                                        FLOOR_BPS,
+                                        no_split,
+                                        use_async,
+                                        subframe_req,
+                                    )
                                     .context(
                                     "NVENC initialize_encoder rejected even at the floor bitrate",
                                 )?;
@@ -1200,6 +1239,7 @@ impl NvencD3d11Encoder {
             self.init_device_com = Some(device.clone());
             // Session init params a later `reconfigure_bitrate` must re-present verbatim.
             self.split_mode = used_split;
+            self.subframe_on = subframe_req;
             self.session_async = use_async;
             // Session-budget accounting (Stage W3): record what this open holds so admission can
             // decline a parallel display the hardware can't afford. Weighted by the FINAL split
@@ -1855,7 +1895,10 @@ impl Encoder for NvencD3d11Encoder {
                     &mut cfg,
                     self.split_mode,
                     self.session_async,
-                    resolve_subframe(false),
+                    // The SESSION's recorded state, not a fresh env read: reconfigure must
+                    // present exactly the init params the open arbitrated (an env re-read could
+                    // flip enableSubFrameWrite mid-session — the Linux latch invariant).
+                    self.subframe_on,
                 ),
                 ..Default::default()
             };

docs-site/content/docs/configuration.md

@@ -143,7 +143,8 @@ notes for context.
 | Setting | Values | Meaning |
 |---|---|---|
 | `PUNKTFUNK_GSO` | `1` · `0` | UDP Generic Segmentation Offload on the send path (coalesce a frame's packets into kernel super-buffers) — cuts send CPU ~30%, but its line-rate packet trains can cost delivered throughput on constrained links (measured on a 2.5GbE hop). Off by default until send pacing spaces the super-buffers; set `1` to opt in (auto-falls back to `sendmmsg` on kernels/paths without support). |
-| `PUNKTFUNK_SPLIT_ENCODE` | `0`/`disable` · `1`/`auto` · `2` · `3` | NVENC N-way split-encode for very high pixel rates (5K@240). `auto` picks automatically above ~1 Gpix/s. |
+| `PUNKTFUNK_SPLIT_ENCODE` | `0`/`disable` · `1`/`auto` · `2` · `3` | NVENC N-way split-encode for very high pixel rates (5K@240). `auto` picks automatically above ~1 Gpix/s. H.264 never splits (not applicable per the SDK); on HEVC a *forced* split disables sub-frame readback (mutually unsupported) — set `0` to choose sub-frame instead. |
+| `PUNKTFUNK_NVENC_SUBFRAME` | `0` · `1` | NVENC sub-frame (slice-level) readback for lower latency on sync sessions. Default: on where the GPU supports it (Linux direct NVENC). `0` = never; `1` = force. On HEVC it yields to a forced split-encode (the SDK documents the pair unsupported). |
 | `PUNKTFUNK_GPU_PRIORITY_CLASS` | `off` · `normal` · `high` · `realtime` · `auto` | **(Windows)** GPU scheduling priority for capture/encode under a GPU-saturating game. Default `auto` (starts `high`, upgrades to `realtime` when it's safe — e.g. HAGS off); `high` pins the static pre-gate behaviour; `realtime` is the strongest lever but can freeze NVENC on some setups. |
 | `PUNKTFUNK_IDD_DEPTH` | `N` (default `2`) | **(Windows)** IDD-push pipeline depth. `1` cuts latency once GPU priority is raised; higher smooths a contended GPU. |