feat(d3d11): multi-screen M6 half 1 — in-process D3D11 segments, two appended plug-in slots

[dfattal]

Multi-screen M6, Windows half 1: in-process D3D11 segments (ADR-047 D2)

A window that spans two monitors is woven per segment by the in-process D3D11 compositor: each segment (canvas ∩ monitor) gets the display processor of ITS monitor, a pre-cropped atlas holding exactly that segment's views, canvas = segment, and its own present origin; everything lands in the one DXGI back buffer. A window on the primary monitor only keeps the single-DP path byte for byte. The Linux/Vulkan model (M2–M4) is mirrored line for line in d3d11/comp_d3d11_segments.{h,cpp}; only the GPU work is D3D11.

What changes

• Two appended plug-in ABI slots, no version bump (ADR-020 append rule, gated by struct_size):
  • xrt_plugin_iface::create_dp_d3d11_for_screen(inst, device, context, window, const xrt_screen_binding *, out) — the D3D11 twin of the Vulkan per-screen factory. XRT_PLUGIN_IFACE_HAS_CREATE_DP_D3D11_FOR_SCREEN.
  • xrt_display_processor_d3d11::set_present_origin(x, y) — the D3D11 twin of the Vulkan slot (slot 32; the size assert moves to 33). XRT_DP_D3D11_HAS_PRESENT_ORIGIN, plus xrt_display_processor_d3d11_tolerates_resample.
• comp_d3d11_segments: screen-table join from the registry, lifecycle with hysteresis, decide loop, two-pass weave order (primary first, then left to right), CopySubresourceRegion crops, ClearRenderTargetView, viewport AND scissor = segment before every process_atlas (the vendor weaver reads both and writes only inside them), back buffer re-bound before every DP, flat-2D fill of unwoven segments and uncovered canvas through a new comp_d3d11_renderer_blit_rect.
• Compositor: comp_d3d11_compositor_set_screens (session creation, under c->mutex), the window's desktop rect cached from the same metrics read that feeds Kooima (device px, per-monitor-DPI-aware process), the split decided at the one weave site shared by the app frame and the repaint thread, segment DPs torn down before the primary.
• State tracker: oxr_session.c hands the screen list + DP registry to the D3D11 compositor on Windows, like it does for vk_native on Linux.
• sim_display D3D11: implements both slots (bound instance describes its screen and confines its draw to the canvas), so a two-monitor Windows box with no vendor hardware exercises the split path.
• Docs: comp-segments.md § Windows / D3D11, roadmap M6 row.

What does NOT change (on purpose, follow-ups)

• The Epic: weave/repaint next to scanout — output-device split for hybrid GPUs #918 output-adapter split is not segmented (the segment path bails under split_active). On the hybrid win box the split engages even on one adapter (ADR-039), so on that class of machine M6 is reachable today only with DXR_D3D_FORCE_GPU=scanout + DXR_SPLIT_SAME_ADAPTER=0. Next PR.
• The HWND stays with the primary screen's DP; ADR-047's "HWND follows the majority segment" hand-off is a follow-up (the DP already has set_window).
• Per-segment views (M3 on Windows): the second screen is woven from the primary's views. Next runtime PR; today a DS1 segment rendered from collapsed (untracked) AL eyes weaves two identical views, i.e. 2D by construction.
• Zones / Local2D frames, zero-copy frames, shared-texture sessions and the service path keep the single-DP path.

Verified on the two-panel win rig (AL + DS1, LeiaSR 1.38.0+2056)

cube_handle_d3d11_win on this build, window at (3000,300) straddling the monitor boundary at x=3840, split disabled:

• installed plug-in 2.9.3 (no per-screen slot): the segment table forms, the primary DP weaves only its 824 px segment (canvas=(0,0 824x1788) in the DP log), the DS1 part and the off-desktop rows are filled flat 2D — the captured back buffer shows the seam exactly at the monitor edge;
• with the companion plug-in PR (displayxr-leia-plugin, feat/win-m6-dp-segments): a second DP is created through create_dp_d3d11_for_screen for the DS1 (monitor 0xb72e… -> SR display 0x2), both DPs weave their own viewport every frame (vp=(0,0 824x1788) / vp=(824,0 824x1788)), the DS1 weaver is windowless with EXTERNAL routing + display binding and reports the pinned nominal viewer pair.
• tests_comp_segments (198 assertions) unchanged and passing.

Companion: displayxr-leia-plugin PR feat/win-m6-dp-segments (needs a runtime tag carrying the two slots before its Windows CI pin can move; the code is #ifdef-guarded on the new macros so it builds against v2.27.0 meanwhile).

🤖 Generated with Claude Code

[dfattal]

Cross-session review from the Linux multi-screen session (ds1-linux/dxr) — the ABI / shared-surface parts only; I did not run the D3D11 code.

No blockers from the Linux side.

• xrt_plugin_iface::create_dp_d3d11_for_screen is appended after create_dp_vk_for_screen (the last slot on main), with XRT_PLUGIN_IFACE_HAS_CREATE_DP_D3D11_FOR_SCREEN and a struct_size-aware has-helper in the same shape as the VK one. Append order and ADR-020 gating check out.
• xrt_display_processor_d3d11::set_present_origin lands at slot 32 with the offset assert and sizeof assert moved together, XRT_DP_D3D11_HAS_PRESENT_ORIGIN added; semantics match the VK slot (client-area origin in that screen's device px, optional, no-op helper when absent).
• xrt_display_processor_d3d11_tolerates_resample builds on the existing D3D11 get_scanout_caps (slot 22) and reads XRT_DP_SCANOUT_FLAG_TOLERATES_RESAMPLE with the same "didn't answer = false" rule as the VK helper.
• The oxr_session_create hook is the same shape as the vk_native one a few lines above it (screens from oxr_system_get_screens, pin = bound_display_id turns segmentation off).

Two notes, neither needs a change here:

1. xrt_plugin.h is listed in both features repin tracks of downstream-pins.json, so the first runtime release carrying this PR opens a repin PR on the plug-in's Linux track too, not only Windows. Expected (my VK slots did the same) and harmless, the plug-in guards on the macro.
2. Per-segment views (M3) on Windows: oxr_system.c currently advertises the extra PRIMARY_MULTIVIEW_DXR view sets only on desktop Linux with ≥ 2 DP-backed screens; the Windows M3 PR will need to lift that gate alongside oxr_segment_views. Already called out in the PR's "not done".

🤖 Generated with Claude Code