feat(d3d11): DXR_WEAVE_GPU_TIMING per-stage weave timing (2D-under-the-lens P0)

[dfattal]

What

P0 of the "near-zero-cost 2D under the lens" plan: instrumentation only, in the D3D11 DP (leia_display_processor_d3d11.cpp). The runtime PR (DisplayXR/displayxr-runtime#1824) extends the service-side DXR_WEAVE_GPU_TIMING=1 diagnostic. Its weave stage covers the whole process_atlas call. This PR splits that stage inside the DP.

Knob

The knob is DXR_WEAVE_GPU_TIMING=1, the same one the runtime's D3D11 service reads. The DP runs in the service process, so one env var arms both lines. DXR_LEIA_DP_WEAVE_TIMING=1 is accepted as a plug-in-only alias.

Both are read with std::getenv and cached, like the plug-in's other knobs. The default is off. When off, the only per-call cost is one cached-bool check; no state is allocated and nothing is logged.

What is measured

GPU. Each 3D process_atlas gets four D3D11 timestamps plus a disjoint query, on the caller's immediate context. They are read back non-blocking from an 8-slot ring, a few frames late. A set still in flight when the ring wraps is dropped, never waited on.

stage	brackets	averaged over
pre_weave	WGC bg compose-under, or the chroma-key fill	frames where a pre-pass ran
sr_weave	2D-layer hand-off (set_compose_filter_strength + set_compose_layer) + leiasr_d3d11_weave, i.e. the SR weave including SR's in-weave compose	all 3D weaves
alpha_gate	the post-weave alpha gate (has_overlay=2 when the layer was composited)	frames where it ran; with 2D layer= counts the has_overlay=2 subset
ck_strip	the chroma-key strip, the other post-weave pass	frames where it ran

CPU. sr_weave: wall time of the leiasr_d3d11_weave call itself, via QPC.

Per-window counts:

• process_atlas calls: 3D weaves, 2D passthrough blits, and not-ready/absent flat blits;
• 2D layer offered: the runtime called set_overlay_2d with a layer;
• 2D layer composed: SR accepted it for this weave, i.e. compose was active;
• 2D layer unchanged: offered and declared unchanged by the client (v14);
• filter strength for composed frames: last, min and max (<0 = SR default);
• bg-compose: WGC compose-under frames.

SR's own refilter count is already in the SR log, so it is not duplicated here.

How to read the log line

One WARN line per DP every 5 s, never per frame:

Leia D3D11 DP timing <window>s: process_atlas=N (3D weave=N, 2D blit=N, not-ready/absent=N) | 2D layer: offered=N composed=N unchanged=N strength last=S min=S max=S (<0 = SR default) | bg-compose=N | CPU sr_weave ms avg/max A/M | GPU ms avg/max n=N disjoint=N: pre_weave A/M (n=N) sr_weave A/M alpha_gate A/M (n=N, with 2D layer=N) ck_strip A/M (n=N)

• The runtime's weave stage ≈ this line's pre_weave + sr_weave + alpha_gate/ck_strip, plus the DP's state-setting overhead.
• composed vs offered: when composed < offered, SR refused the layer at weave time (it is also logged once as an error).
• unchanged/offered is how often the client let SR reuse its prefilter. Compare it with SR's refilter count in the SR log.
• A large gap between CPU sr_weave and GPU sr_weave means SR's CPU-side work (predictor, state), not GPU work.

Build

Built with scripts\build-windows.bat build from a worktree, against the local ../displayxr-runtime headers (ABI 5): green (DisplayXR-LeiaSR.dll linked and staged to _package/bin/plugins/). No runtime header changes are needed; this PR is independent of the runtime PR at compile time.

Testing

Not panel-tested yet; the hub session runs the panel A/B. Deploy by rename-swapping the dev DLL into C:\Program Files\DisplayXR\Plugins\LeiaSR\, then start the service with DXR_WEAVE_GPU_TIMING=1 in its environment.

🤖 Generated with Claude Code