Implement render-first pacing for Vulkan backend and enhance view location handling

OPENXR.md

@@ -320,6 +320,19 @@ short-lived immutable stereo packet. Each sealed GX frame and immersive packet c
 policy-generation tag; a mismatch is rendered in mono and the acquired XR frame is canceled, so an
 asynchronous menu/race transition cannot replay race transforms over unsafe content.
 
+With interpolation off, the standalone (Vulkan) backend paces render-first: the pacing thread
+locates the views for an estimated display time (two periods past the last one the compositor
+predicted), hands Aurora the packet and its shared-buffer targets with no compositor frame open,
+and waits for Aurora to render the eyes at its next seal, repeating the retained layer if that
+takes more than 50 ms. Only then does it call xrWaitFrame and xrBeginFrame, copy the eyes into the
+freshly acquired swapchain images and end the frame with the packet's render poses. A headset
+frame therefore stays open for the copy alone instead of for the next 60 Hz game frame plus the
+whole encode, which on a 72 or 90 Hz display used to make every second cycle span two display
+slots (about 45 headset frames per second while the game rendered 60). The compositor reprojects
+the rendered pose to the frame it lands in. The D3D12 backend keeps the frame-first order below,
+as does interpolation on either backend, since interpolation renders for the frame's own
+predicted display time.
+
 With VR interpolation enabled, Aurora retains each sealed race's command stream and matched
 previous/current transform uniforms. New OpenXR packets wake the frame worker between game
 frames. It interpolates at the requested display time, then applies that packet's head pose and

docs/quest-port.md

@@ -661,12 +661,31 @@ stereo frame cost 13.2 ms, of which the native render was 5.7 ms, the eyes 3.5
 and 3.8, copies and gaps 0.4. That native render is a 1280x720 image nobody
 sees during an immersive race, so it now stops after the last pass whose EFB
 copy the eyes sample: `mono` fell to 0.15 ms and a Luigi Circuit start at 0.5
-renders in 5.5 to 10 ms of GPU per frame. What remains is the headset pacing:
-with the display at 72 or 90 Hz, each headset frame stays open for the next
-60 Hz game frame plus the whole encode (`open` 16 ms in the pacing summary),
-so cycles span one to two display slots and the headset gets 40 to 60 frames
-per second while the game renders 60. Reworking that pacing (encode the newest
-sealed frame at once, repeat the layer otherwise) is the next step.
+renders in 5.5 to 10 ms of GPU per frame. The last limiter was the headset
+pacing: with the display at 72 or 90 Hz, each headset frame stayed open for
+the next 60 Hz game frame plus the whole encode (`open` 16 ms in the pacing
+summary), so cycles spanned one to two display slots and the headset got 40 to
+60 frames per second while the game rendered 60. The Vulkan backend now paces
+render-first (`PreparePacket`, `BeginFrameForPacket`, `CopyRenderedEyes` in
+`openxr_vulkan.cpp`; see `OPENXR.md`): the packet is located and handed to
+Aurora with no compositor frame open, and the frame is begun only once the
+eyes exist, for the copy alone. On the same automated start at 0.75 the
+summary reads `cycles=60 skipped-slots=12 late=0 layers new=60 repeat=0
+open=5.5 end-gap=16.7`, the compositor shows 60 to 61 of 72 with the
+inherent 12 stale slots, app-to-compositor latency fell from 51 to 9 to 13 ms,
+and the frame worker's encode fell from 8 to 2.7 ms because the eye copy and
+its fence wait moved off the worker onto the pacing thread.
+
+Retro Rewind tracks then showed a game-thread limit of their own: on Athens
+Dash (a Mario Kart Tour port) the display-list index scan
+(`WalkDisplayList<DlIndexScanVisitor>`) was 11.5% of the thread while the base
+game's tracks spend 0.3% there. The scan cache in `gx_dl.cpp` refused lists
+above 64 KiB, so that track's large shape lists were scanned again on every
+call; the cap is now 4 MiB. With it the scan is 0.2%, the game rate on Athens
+Dash went from 47 to 51 fps to 50 to 58, and the thread splits into 62% game
+plus mod code, 9% GX HLE, 6% FIFO decode, 4% memory copies, 3.5% dispatch and
+the rest. What remains on such tracks is the game's own code plus the mod's,
+which no host change shrinks; a GX thread could move about 20% of it.
 
 Verified on device since: the menus on the virtual screen, controller input
 (the user has driven races), and an immersive Grand Prix start with all 12

runtime/include/vr/openxr_runtime.h

@@ -153,6 +153,11 @@ public:
     OpenXRFrameStatus WaitFrame(OpenXRFrame& frame);
     bool BeginFrame(const OpenXRFrame& frame);
     bool LocateViews(OpenXRFrame& frame);
+    // Locates the views for `display_time` outside the frame protocol, for a
+    // packet whose eyes are rendered before the compositor frame that will show
+    // them is begun (the standalone backend's render-first pacing). Fills the
+    // frame's display time, views, flags and validity; requires a running session.
+    bool LocateViewsAt(XrTime display_time, OpenXRFrame& frame);
     bool EndFrame(
         const OpenXRFrame& frame,
         const XrCompositionLayerBaseHeader* const* layers,
@@ -224,6 +229,7 @@ private:
     bool EnumerateSwapchainFormats();
     bool HandleSessionStateChanged(const XrEventDataSessionStateChanged& event);
     bool IsFrameTokenCurrent(const OpenXRFrame& frame, FramePhase expected) const;
+    bool LocateViewsForFrame(OpenXRFrame& frame);
     void ResetFrameState();
     void DestroyReferenceSpaces();
     void ResetSessionState();

runtime/include/vr/openxr_vulkan.h

@@ -58,6 +58,20 @@ public:
     bool RepeatFrame(const OpenXRBackendFrame& frame);
     bool FinishFrame(OpenXRBackendFrame& frame, bool submit_layer);
 
+    // Render-first pacing, used while VR interpolation is off. A packet is prepared with the
+    // views located for an estimated display time and handed to Aurora without a compositor
+    // frame open; once Aurora has rendered the eyes into the shared buffers, the compositor frame
+    // is begun, the eyes are copied into its swapchain images and it is ended at once. A headset
+    // frame therefore never waits for a game frame: it stays open for the copy alone.
+    OpenXRBeginStatus PreparePacket(const OpenXRPresentation& presentation, OpenXRBackendFrame& packet);
+    bool TryCancelPendingPacket(OpenXRBackendFrame& packet);
+    OpenXRBeginStatus BeginFrameForPacket(const OpenXRBackendFrame& packet, OpenXRBackendFrame& frame);
+    OpenXRSubmissionStatus CopyRenderedEyes(const OpenXRBackendFrame& frame);
+    // One compositor cycle that resubmits the retained layer (or nothing), with no frame left
+    // active: keeps the runtime fed while the eyes are still being rendered and learns the
+    // display timing the next packet is located for.
+    OpenXRBeginStatus KeepAliveCycle();
+
     // Drains this backend's own queue before tearing down. Returns false only
     // when the private device could not be waited on, in which case the caller
     // retains the backend and runtime for the process lifetime.

runtime/src/hle/gx/gx_dl.cpp

@@ -28,9 +28,13 @@ using GxCpDecode::SameVtxAttrFmt;
 // Small display lists dominate the in-race call count. Cache them as well, but
 // cap both individual entries and aggregate copied command bytes so malformed
 // guest input cannot turn this optimization into unbounded host allocation.
-constexpr uint32_t kDlScanCacheMaxEntryBytes = 64u * 1024u;
+// The entry cap was 64 KiB: a list above it was never cached and its index
+// scan ran on every call, which on a Retro Rewind track with large shape lists
+// was 11% of the game thread on a Quest 3. Only lists that need flattening
+// store a copy, and the aggregate cap still bounds those.
+constexpr uint32_t kDlScanCacheMaxEntryBytes = 4u * 1024u * 1024u;
 constexpr size_t kDlScanCacheMaxEntries = 8192;
-constexpr size_t kDlScanCacheMaxStoredBytes = 8u * 1024u * 1024u;
+constexpr size_t kDlScanCacheMaxStoredBytes = 32u * 1024u * 1024u;
 
 // Display-list write tracking (audit F6a): re-digesting every list every call is the
 // costliest step of GX__CallDisplayList, and wasted on BRRES shape lists that are written

runtime/src/vr/openxr_integration.cpp

@@ -802,6 +802,19 @@ private:
             const uint32_t interpolation_target = frame_interpolation_fps_.load(std::memory_order_relaxed);
             SetInterpolationActive(immersive && FrameInterpolationAvailable() && interpolation_target != 0);
 
+#if !defined(_WIN32)
+            // Standalone (Vulkan) backend: with interpolation off, the eyes are rendered before
+            // the compositor frame that shows them is begun, so that frame never waits for a
+            // game frame. Interpolation keeps the frame-first order below: it renders for the
+            // frame's own predicted display time.
+            if (!aurora_get_stereo_frame_interpolation()) {
+                if (!RenderFirstCycle(presentation, policy, immersive, consecutive_skips,
+                                      immersive_submission_logged)) {
+                    fatal = true;
+                }
+                continue;
+            }
+#endif
             OpenXRBackendFrame frame{};
             const OpenXRBeginStatus begin = backend_->BeginFrame(presentation, frame);
             if (begin == OpenXRBeginStatus::SessionNotRunning) {
@@ -961,6 +974,164 @@ private:
         ShutdownOrRetainGraphicsObjects();
     }
 
+#if !defined(_WIN32)
+    // One compositor cycle on the retained layer, with no frame left active. False on a fatal
+    // backend or runtime failure (the error is recorded).
+    bool KeepAlive() {
+        const OpenXRBeginStatus status = backend_->KeepAliveCycle();
+        if (status == OpenXRBeginStatus::SessionNotRunning) {
+            MkwVRPolicySetSessionActive(false);
+            return true;
+        }
+        if (status == OpenXRBeginStatus::ExitRequested) {
+            SetError("OpenXR runtime requested session exit; continuing on the mirror output");
+            return false;
+        }
+        if (status == OpenXRBeginStatus::Error) {
+            SetError(backend_->LastError());
+            return false;
+        }
+        return true;
+    }
+
+    // Render-first pacing (see OpenXRVulkanBackend::PreparePacket). Returns false on a fatal
+    // failure; a cycle that ends without a layer returns true and the loop tries again.
+    bool RenderFirstCycle(OpenXRPresentation presentation, const MkwVRPolicySnapshot& policy, bool immersive,
+                          uint32_t& consecutive_skips, bool& immersive_submission_logged) {
+        OpenXRBackendFrame packet{};
+        const OpenXRBeginStatus prepared = backend_->PreparePacket(presentation, packet);
+        if (prepared == OpenXRBeginStatus::SessionNotRunning) {
+            MkwVRPolicySetSessionActive(false);
+            return true;
+        }
+        if (prepared == OpenXRBeginStatus::ExitRequested) {
+            SetError("OpenXR runtime requested session exit; continuing on the mirror output");
+            return false;
+        }
+        if (prepared == OpenXRBeginStatus::Error) {
+            SetError(backend_->LastError());
+            return false;
+        }
+        // The head pose this packet was located with places the screens and aims the pointer.
+        ServiceRecenterRequest();
+        UpdateVirtualScreenPose(packet);
+        if (input_ != nullptr) {
+            input_->Sync(packet.xr_frame.predicted_display_time, PointerScreen(packet, policy, immersive),
+                         SettingsPanelScreen(packet, policy, immersive));
+        }
+        if (!packet.expects_gpu_submission) {
+            // Nothing to render (no rendering requested or no tracking): keep the compositor fed.
+            return KeepAlive();
+        }
+        {
+            std::lock_guard lock(published_mutex_);
+            BuildPublishedFrame(packet, immersive, policy.EffectiveUnitsPerMeter(), policy.content_tag);
+            diagnostics::OnPacketPublished();
+            published_.store(&published_frame_, std::memory_order_release);
+        }
+        aurora_notify_stereo_frame();
+
+        // Aurora renders the eyes at its next seal. Meanwhile the compositor keeps showing the
+        // retained layer; a 50 ms stall repeats it explicitly and withdraws the packet.
+        OpenXRSubmissionStatus submission = OpenXRSubmissionStatus::Timeout;
+        bool canceled_before_encode = false;
+        const auto cancel_after = std::chrono::steady_clock::now() + std::chrono::milliseconds(50);
+        while (!stop_.load(std::memory_order_acquire) && submission == OpenXRSubmissionStatus::Timeout) {
+            submission = backend_->WaitForSubmission(packet, 50);
+            if (submission == OpenXRSubmissionStatus::Timeout) {
+                if (std::chrono::steady_clock::now() >= cancel_after) {
+                    WithdrawPublishedFrame();
+                    canceled_before_encode = backend_->TryCancelPendingPacket(packet);
+                    if (canceled_before_encode) {
+                        diagnostics::OnPacketCanceled();
+                        break;
+                    }
+                }
+                diagnostics::OnKeepaliveRepeat();
+                if (!KeepAlive()) {
+                    return false;
+                }
+            }
+        }
+        WithdrawPublishedFrame();
+        if (stop_.load(std::memory_order_acquire) || canceled_before_encode ||
+            submission == OpenXRSubmissionStatus::ShuttingDown) {
+            return true;
+        }
+        if (submission != OpenXRSubmissionStatus::Success) {
+            // Nothing reached the shared buffers (Skipped) or Aurora failed after queuing GPU
+            // work (Failed): same accounting as the frame-first path, on a keep-alive cycle.
+            diagnostics::OnSubmission(false);
+            if (submission == OpenXRSubmissionStatus::Failed) {
+                SetError(std::string("Aurora's ") + kGraphicsBackendName +
+                         " stereo copy failed; continuing on the mirror output");
+                return false;
+            }
+            ++consecutive_skips;
+            if (consecutive_skips == 1 || consecutive_skips % 60 == 0) {
+                RT_LOG(RT_TAG_RUNTIME) << "OpenXR: eye copy skipped (" << consecutive_skips
+                                       << " in a row): " << backend_->LastError() << std::endl;
+            }
+            if (consecutive_skips >= kMaxConsecutiveSkips) {
+                SetError(std::string("Aurora's ") + kGraphicsBackendName +
+                         " stereo copy keeps failing; continuing on the mirror output");
+                return false;
+            }
+            return KeepAlive();
+        }
+
+        // The eyes are in the shared buffers: begin the compositor frame, copy, end.
+        OpenXRBackendFrame frame{};
+        const OpenXRBeginStatus begin = backend_->BeginFrameForPacket(packet, frame);
+        if (begin == OpenXRBeginStatus::SessionNotRunning) {
+            MkwVRPolicySetSessionActive(false);
+            return true;
+        }
+        if (begin == OpenXRBeginStatus::ExitRequested) {
+            SetError("OpenXR runtime requested session exit; continuing on the mirror output");
+            return false;
+        }
+        if (begin == OpenXRBeginStatus::Error) {
+            SetError(backend_->LastError());
+            return false;
+        }
+        UpdateFrameTiming(frame.xr_frame);
+        if (diagnostics::Enabled()) {
+            NoteFrameDiagnostics(frame, immersive);
+        }
+        const OpenXRSubmissionStatus copy =
+            frame.expects_gpu_submission ? backend_->CopyRenderedEyes(frame) : OpenXRSubmissionStatus::Skipped;
+        const bool submit = copy == OpenXRSubmissionStatus::Success;
+        diagnostics::OnSubmission(submit);
+        if (!backend_->FinishFrame(frame, submit)) {
+            SetError(backend_->LastError());
+            return false;
+        }
+        if (copy == OpenXRSubmissionStatus::Failed) {
+            SetError(std::string("Aurora's ") + kGraphicsBackendName +
+                     " stereo copy failed; continuing on the mirror output");
+            return false;
+        }
+        if (!submit) {
+            ++consecutive_skips;
+            if (consecutive_skips == 1 || consecutive_skips % 60 == 0) {
+                RT_LOG(RT_TAG_RUNTIME) << "OpenXR: eye copy skipped (" << consecutive_skips
+                                       << " in a row): " << backend_->LastError() << std::endl;
+            }
+            return consecutive_skips < kMaxConsecutiveSkips;
+        }
+        consecutive_skips = 0;
+        ++timing_submissions_;
+        if (immersive && !immersive_submission_logged) {
+            immersive_submission_logged = true;
+            RT_LOG(RT_TAG_RUNTIME) << "[mkw-vr] first immersive packet consumed and submitted as "
+                                      "an OpenXR projection layer"
+                                   << std::endl;
+        }
+        return true;
+    }
+#endif
+
     void BuildPublishedFrame(const OpenXRBackendFrame& source, bool immersive,
                              float units_per_meter, uint64_t content_tag) noexcept {
         ApplyPendingReferenceSpaceChange(source.xr_frame);

runtime/src/vr/openxr_runtime.cpp

@@ -741,6 +741,19 @@ bool OpenXRRuntime::LocateViews(OpenXRFrame& frame) {
         return Fail(XR_ERROR_CALL_ORDER_INVALID, "xrLocateViews",
                     "frame token is stale or xrBeginFrame was not called");
     }
+    return LocateViewsForFrame(frame);
+}
+
+bool OpenXRRuntime::LocateViewsAt(XrTime display_time, OpenXRFrame& frame) {
+    ClearError();
+    if (!HasSession() || !m_session_running) {
+        return Fail(XR_ERROR_SESSION_NOT_RUNNING, "xrLocateViews", "the session is not running");
+    }
+    frame.predicted_display_time = display_time;
+    return LocateViewsForFrame(frame);
+}
+
+bool OpenXRRuntime::LocateViewsForFrame(OpenXRFrame& frame) {
     frame.views_valid = false;
     frame.view_state_flags = 0;
     if (!frame.should_render) {

runtime/src/vr/openxr_vulkan.cpp

@@ -482,6 +482,139 @@ public:
         return true;
     }
 
+    // ---- Render-first pacing (see openxr_vulkan.h) ---------------------------------------------
+
+    void DiscardPendingReleasesLocked() noexcept {
+        if (!have_pending_releases_) {
+            return;
+        }
+        for (auto& release : pending_releases_) {
+            CloseFd(release.releaseFenceFd);
+            release = {-1, VK_IMAGE_LAYOUT_UNDEFINED};
+        }
+        have_pending_releases_ = false;
+    }
+
+    OpenXRBeginStatus PreparePacket(const OpenXRPresentation& presentation, OpenXRBackendFrame& packet) {
+        packet = {};
+        packet.presentation = presentation;
+        if (!bound_ || runtime_ == nullptr) {
+            Fail("PreparePacket called before the Vulkan backend was bound");
+            return OpenXRBeginStatus::Error;
+        }
+        if (frame_active_) {
+            Fail("PreparePacket called while an OpenXR frame is active");
+            return OpenXRBeginStatus::Error;
+        }
+        if (runtime_->ShouldExit()) {
+            return OpenXRBeginStatus::ExitRequested;
+        }
+        if (!runtime_->IsSessionRunning()) {
+            return OpenXRBeginStatus::SessionNotRunning;
+        }
+        if (last_display_period_ <= 0) {
+            // No display timing yet: one compositor cycle learns it.
+            const OpenXRBeginStatus primed = KeepAliveCycle();
+            if (primed != OpenXRBeginStatus::Ready) {
+                return primed;
+            }
+        }
+        packet.xr_frame.serial = next_packet_serial_++;
+        // The eyes are ready after at most one game frame plus the encode and show at the first
+        // display slot after that: two periods past the last predicted display time.
+        packet.xr_frame.predicted_display_time = last_display_time_ + 2 * last_display_period_;
+        packet.xr_frame.predicted_display_period = last_display_period_;
+        packet.xr_frame.should_render = last_should_render_;
+        for (uint32_t eye = 0; eye < kOpenXREyeCount; ++eye) {
+            packet.render_width[eye] = eye_swapchains_[eye].width;
+            packet.render_height[eye] = eye_swapchains_[eye].height;
+        }
+        if (!packet.xr_frame.should_render) {
+            return OpenXRBeginStatus::Ready;
+        }
+        if (!runtime_->LocateViewsAt(packet.xr_frame.predicted_display_time, packet.xr_frame)) {
+            Fail("xrLocateViews failed for a packet");
+            return OpenXRBeginStatus::Error;
+        }
+        if (!packet.xr_frame.views_valid) {
+            return OpenXRBeginStatus::Ready;
+        }
+
+        const uint32_t target_count =
+            presentation.mode == OpenXRFrameMode::VirtualScreen ? 1u : kOpenXREyeCount;
+        if (target_count == 1) {
+            packet.render_width[1] = packet.render_width[0];
+            packet.render_height[1] = packet.render_height[0];
+        }
+        std::array<AuroraVulkanStereoTarget, kOpenXREyeCount> targets{};
+        const uint32_t slot = next_slot_;
+        {
+            std::lock_guard lock(vk_mutex_);
+            DiscardPendingReleasesLocked();
+            for (uint32_t eye = 0; eye < target_count; ++eye) {
+                EyeSlot& eye_slot = slots_[eye][slot];
+                targets[eye] = {
+                    eye_slot.buffer,
+                    eye_slot.width,
+                    eye_slot.height,
+                    static_cast<int64_t>(aurora_format_),
+                    DupFd(eye_slot.pending_acquire_fd),
+                    static_cast<int32_t>(eye_slot.layout),
+                };
+            }
+            // The compositor images are acquired by BeginFrameForPacket, once the eyes exist.
+            pending_copy_ = {packet.xr_frame.serial, slot, target_count, {}};
+            deferred_copy_ = true;
+        }
+        {
+            std::lock_guard lock(submission_mutex_);
+            awaiting_token_ = packet.xr_frame.serial;
+            submitted_token_ = 0;
+            submission_arrived_ = false;
+            submission_success_ = false;
+            submission_unsafe_ = false;
+        }
+        if (!aurora_vulkan_set_stereo_targets(packet.xr_frame.serial, targets.data(), target_count)) {
+            for (uint32_t eye = 0; eye < target_count; ++eye) {
+                CloseFd(targets[eye].acquireFenceFd);
+            }
+            {
+                std::lock_guard lock(submission_mutex_);
+                awaiting_token_ = 0;
+            }
+            {
+                std::lock_guard lock(vk_mutex_);
+                deferred_copy_ = false;
+            }
+            Fail("Aurora rejected the AHardwareBuffer stereo targets");
+            return OpenXRBeginStatus::Error;
+        }
+        next_slot_ = (slot + 1) % kSlotCount;
+        packet.expects_gpu_submission = true;
+        return OpenXRBeginStatus::Ready;
+    }
+
+    bool TryCancelPendingPacket(OpenXRBackendFrame& packet) {
+        if (!packet.expects_gpu_submission || !aurora_vulkan_cancel_stereo_targets(packet.xr_frame.serial)) {
+            return false;
+        }
+        {
+            std::lock_guard lock(submission_mutex_);
+            awaiting_token_ = 0;
+            submitted_token_ = 0;
+            submission_arrived_ = false;
+            submission_success_ = false;
+            submission_unsafe_ = false;
+        }
+        {
+            std::lock_guard lock(vk_mutex_);
+            DiscardPendingReleasesLocked();
+            deferred_copy_ = false;
+        }
+        packet.expects_gpu_submission = false;
+        return true;
+    }
+
     bool FinishFrame(OpenXRBackendFrame& frame, bool submit_layer) {
         if (!frame_active_ || runtime_ == nullptr ||
             frame.xr_frame.serial != active_frame_serial_) {
@@ -531,6 +664,12 @@ public:
         active_frame_serial_ = 0;
         active_frame_ = {};
         frame.expects_gpu_submission = false;
+        {
+            // Eyes rendered for a packet that this frame did not copy are dropped with it.
+            std::lock_guard lock(vk_mutex_);
+            DiscardPendingReleasesLocked();
+            deferred_copy_ = false;
+        }
         {
             std::lock_guard lock(submission_mutex_);
             awaiting_token_ = 0;
@@ -541,6 +680,137 @@ public:
         return release_ok && end_ok;
     }
 
+    OpenXRBeginStatus BeginFrameForPacket(const OpenXRBackendFrame& packet, OpenXRBackendFrame& frame) {
+        frame = {};
+        frame.presentation = packet.presentation;
+        frame.render_width = packet.render_width;
+        frame.render_height = packet.render_height;
+        if (!bound_ || runtime_ == nullptr) {
+            Fail("BeginFrameForPacket called before the Vulkan backend was bound");
+            return OpenXRBeginStatus::Error;
+        }
+        if (frame_active_) {
+            Fail("BeginFrameForPacket called while another OpenXR frame is active");
+            return OpenXRBeginStatus::Error;
+        }
+        const OpenXRFrameStatus status = runtime_->WaitFrame(frame.xr_frame);
+        if (status != OpenXRFrameStatus::Ready) {
+            if (status == OpenXRFrameStatus::Error) {
+                Fail(BeginStatusOperation(status));
+            }
+            return status == OpenXRFrameStatus::SessionNotRunning ? OpenXRBeginStatus::SessionNotRunning
+                   : status == OpenXRFrameStatus::ExitRequested  ? OpenXRBeginStatus::ExitRequested
+                                                                  : OpenXRBeginStatus::Error;
+        }
+        NoteDisplayTiming(frame.xr_frame);
+        if (!runtime_->BeginFrame(frame.xr_frame)) {
+            Fail("xrBeginFrame failed");
+            return OpenXRBeginStatus::Error;
+        }
+        frame_active_ = true;
+        active_frame_serial_ = frame.xr_frame.serial;
+        render_session_serial_ = runtime_->SessionRunSerial();
+        render_space_serial_ = runtime_->LastReferenceSpaceChange().serial;
+        // The layer shows the eyes as they were rendered: it carries the packet's located views.
+        frame.xr_frame.views = packet.xr_frame.views;
+        frame.xr_frame.view_state_flags = packet.xr_frame.view_state_flags;
+        frame.xr_frame.views_valid = packet.xr_frame.views_valid;
+        active_frame_ = frame.xr_frame;
+        frame.expects_gpu_submission = packet.expects_gpu_submission;
+        if (!frame.xr_frame.should_render || !frame.expects_gpu_submission || !frame.xr_frame.views_valid) {
+            // No swapchain image is acquired for this frame, so the rendered eyes cannot be
+            // copied; FinishFrame drops them with the frame.
+            frame.expects_gpu_submission = false;
+            return OpenXRBeginStatus::Ready;
+        }
+
+        const uint32_t target_count =
+            presentation_target_count(frame.presentation);
+        const diagnostics::Stopwatch acquire_timer;
+        for (uint32_t eye = 0; eye < target_count; ++eye) {
+            if (!AcquireSwapchain(eye_swapchains_[eye])) {
+                ReleaseAcquiredSwapchains();
+                EndActiveFrameWithoutLayers(frame.xr_frame);
+                return OpenXRBeginStatus::Error;
+            }
+        }
+        diagnostics::OnSwapchainAcquire(acquire_timer);
+        {
+            std::lock_guard lock(vk_mutex_);
+            for (uint32_t eye = 0; eye < target_count; ++eye) {
+                pending_copy_.swapchain_images[eye] =
+                    eye_swapchains_[eye].images[eye_swapchains_[eye].acquired_index].image;
+            }
+        }
+        return OpenXRBeginStatus::Ready;
+    }
+
+    OpenXRSubmissionStatus CopyRenderedEyes(const OpenXRBackendFrame& frame) {
+        CopyOutcome outcome = CopyOutcome::Skipped;
+        {
+            std::lock_guard lock(vk_mutex_);
+            if (!frame_active_ || !have_pending_releases_ || !deferred_copy_) {
+                return OpenXRSubmissionStatus::Skipped;
+            }
+            outcome = RecordAndSubmitCopyLocked(pending_releases_);
+            have_pending_releases_ = false;
+            deferred_copy_ = false;
+        }
+        {
+            // FinishFrame judges the queue's safety by this frame's token.
+            std::lock_guard lock(submission_mutex_);
+            submitted_token_ = frame.xr_frame.serial;
+            submission_arrived_ = true;
+            submission_success_ = outcome == CopyOutcome::Submitted;
+            submission_unsafe_ = outcome == CopyOutcome::Unsafe;
+        }
+        return outcome == CopyOutcome::Submitted ? OpenXRSubmissionStatus::Success
+               : outcome == CopyOutcome::Unsafe  ? OpenXRSubmissionStatus::Failed
+                                                 : OpenXRSubmissionStatus::Skipped;
+    }
+
+    OpenXRBeginStatus KeepAliveCycle() {
+        if (!bound_ || runtime_ == nullptr || frame_active_) {
+            Fail("KeepAliveCycle called with a frame active or before binding");
+            return OpenXRBeginStatus::Error;
+        }
+        OpenXRFrame cycle{};
+        const OpenXRFrameStatus status = runtime_->WaitFrame(cycle);
+        if (status != OpenXRFrameStatus::Ready) {
+            if (status == OpenXRFrameStatus::Error) {
+                Fail(BeginStatusOperation(status));
+            }
+            return status == OpenXRFrameStatus::SessionNotRunning ? OpenXRBeginStatus::SessionNotRunning
+                   : status == OpenXRFrameStatus::ExitRequested  ? OpenXRBeginStatus::ExitRequested
+                                                                  : OpenXRBeginStatus::Error;
+        }
+        NoteDisplayTiming(cycle);
+        if (!runtime_->BeginFrame(cycle)) {
+            Fail("xrBeginFrame failed for a keep-alive cycle");
+            return OpenXRBeginStatus::Error;
+        }
+        active_frame_ = cycle;
+        frame_active_ = true;
+        const bool end_ok = EndRetainedFrame(false);
+        frame_active_ = false;
+        active_frame_ = {};
+        if (!end_ok) {
+            Fail("OpenXR could not resubmit the retained frame");
+            return OpenXRBeginStatus::Error;
+        }
+        return OpenXRBeginStatus::Ready;
+    }
+
+    void NoteDisplayTiming(const OpenXRFrame& frame) noexcept {
+        last_display_time_ = frame.predicted_display_time;
+        last_display_period_ = frame.predicted_display_period;
+        last_should_render_ = frame.should_render;
+    }
+
+    static uint32_t presentation_target_count(const OpenXRPresentation& presentation) noexcept {
+        return presentation.mode == OpenXRFrameMode::VirtualScreen ? 1u : kOpenXREyeCount;
+    }
+
     bool RepeatFrame(const OpenXRBackendFrame& frame) {
         if (!frame_active_ || runtime_ == nullptr ||
             frame.xr_frame.serial != active_frame_serial_) {
@@ -1082,6 +1352,7 @@ private:
 
     void DestroySlots() {
         std::lock_guard lock(vk_mutex_);
+        DiscardPendingReleasesLocked();
         if (vk_device_ != VK_NULL_HANDLE) {
             vkDeviceWaitIdle(vk_device_);
         }
@@ -1136,7 +1407,16 @@ private:
                 expected = token == self->awaiting_token_ && token == self->pending_copy_.token;
             }
             if (expected && success && release_count >= self->pending_copy_.target_count) {
+                if (self->deferred_copy_) {
+                    // No compositor frame is open yet: keep Dawn's release fences for
+                    // CopyRenderedEyes, which records the copy once the frame is begun.
+                    self->DiscardPendingReleasesLocked();
+                    self->pending_releases_ = owned;
+                    self->have_pending_releases_ = true;
+                    outcome = CopyOutcome::Submitted;
+                } else {
                     outcome = self->RecordAndSubmitCopyLocked(owned);
+                }
             } else {
                 for (auto& release : owned) {
                     CloseFd(release.releaseFenceFd);
@@ -1594,6 +1874,15 @@ private:
     uint32_t next_submission_ = 0;
     uint32_t next_slot_ = 0;
     PendingCopy pending_copy_{};
+    // Render-first pacing (PreparePacket): Aurora's release fences arrive while no compositor
+    // frame is active, so the copy is recorded later by CopyRenderedEyes.
+    bool deferred_copy_ = false;
+    std::array<AuroraVulkanStereoRelease, kOpenXREyeCount> pending_releases_{};
+    bool have_pending_releases_ = false;
+    uint64_t next_packet_serial_ = 1ull << 40; // never collides with the runtime's frame serials
+    XrTime last_display_time_ = 0;
+    XrDuration last_display_period_ = 0;
+    bool last_should_render_ = false;
 
     std::mutex submission_mutex_;
     std::condition_variable submission_cv_;
@@ -1638,6 +1927,21 @@ OpenXRSubmissionStatus OpenXRVulkanBackend::WaitForSubmission(const OpenXRBacken
                                                               uint32_t timeout_ms) {
     return m_impl->WaitForSubmission(frame, timeout_ms);
 }
+OpenXRBeginStatus OpenXRVulkanBackend::PreparePacket(const OpenXRPresentation& presentation,
+                                                     OpenXRBackendFrame& packet) {
+    return m_impl->PreparePacket(presentation, packet);
+}
+bool OpenXRVulkanBackend::TryCancelPendingPacket(OpenXRBackendFrame& packet) {
+    return m_impl->TryCancelPendingPacket(packet);
+}
+OpenXRBeginStatus OpenXRVulkanBackend::BeginFrameForPacket(const OpenXRBackendFrame& packet,
+                                                           OpenXRBackendFrame& frame) {
+    return m_impl->BeginFrameForPacket(packet, frame);
+}
+OpenXRSubmissionStatus OpenXRVulkanBackend::CopyRenderedEyes(const OpenXRBackendFrame& frame) {
+    return m_impl->CopyRenderedEyes(frame);
+}
+OpenXRBeginStatus OpenXRVulkanBackend::KeepAliveCycle() { return m_impl->KeepAliveCycle(); }
 
 bool OpenXRVulkanBackend::TryCancelPendingFrame(OpenXRBackendFrame& frame) {
     return m_impl->TryCancelPendingFrame(frame);