Opt-in renderer cost profile

K230_VGLITE_PROFILE=1 records wall and process CPU durations for each render pass. The ordinary image remains unchanged; other values disable profiling. The root diagnostic harness accepts --profile and passes that setting only to its compositor. It retains its unprivileged client and recovery watchdog.

Phases cover immutable snapshot creation, imported-target cache maintenance, VG-Lite initialization, target import, source allocation/conversion, GPU API commands (including SDK source cache cleaning), synchronous finish, GPU resource cleanup, and full Pixman replay. Total spans pass creation through source release. It includes other pass bookkeeping and diagnostic logging; phase sums need not equal total. Setup/clock/log overhead and capture disturbance remain. Process CPU includes user/system time charged to the process; it excludes other processes, separately accounted interrupts and whole-device cost. No optical latency or scanout cadence is measured by these timers.

The production renderer sanitizer suite passes with profiling both disabled and enabled, retaining exact pinned Pixman comparisons and failure quarantine. Eight root-harness tests pass, including compositor-only opt-in, as do four parser tests rejecting absent, malformed, failed, insufficient or mixed runs. The narrow nix build .#shell-compositor-vglite --max-jobs 1 --cores 8 --no-link --print-out-paths passes and yields /nix/store/kb31nr0jgpz8bcpzy26s7l50r99vzxib-sway-1.12.

TMPDIR=/mnt/MediaVolume/home/jadams \
WLROOTS_SOURCE=/nix/store/r2zcb3d3dgmz09qjaqsy945inal2h41r-source \
VGLITE_SOURCE=/nix/store/pvpqrg3diyi9mcgmqjc0f80hpdmk699p-source/buildroot-overlay/package/vg_lite \
  tools/test-vglite-renderer.sh
python3 tests/vglite/test_root_trial.py
python3 tests/vglite/test_cost.py
python3 tools/vglite-cost.py COST_LOG --minimum-frames 100 --discard-frames 20

Board profiling is recorded below. This is a diagnostic for the next optimization decision, not completion of the matched performance, normal-service, cache or physical-control acceptance gates.

First measured board pair

The source-built artifact above ran the same animated parent/child fixture on native 568x1232 RGB565 scanout. scene-check.py records the exact harness and client invocation. Both runs returned zero, observed one Sway process, and restored the normal shell and seatd on the unchanged boot. The GPU run completed 182 GPU frames with no replays; forced Pixman completed 571 replays. After excluding 20 warmup frames, the parser retained 162 GPU and 551 CPU samples. baseline/*/costs.log contains every original fixed-schema sample; matching summary.json files record the command’s aggregates and limits. Decision logs and actual native captures are retained too. The same palette checker still passes exactly for parent and child in both captures/modes.

Mean render-pass costs in milliseconds:

This identifies per-frame GPU context lifetime as the first optimization target. The renderer currently calls vg_lite_init and vg_lite_close for each pass. The pinned running kernel source (Xuantie revision 7d4e1f444f461dbe3833bd99a4640e7b6c2cd529), drivers/gpu/vglite/vg_lite_kernel.c:gpu/init_vglite/terminate_vglite, switches clocks and resets hardware on the context-reference transitions 0→1 and 1→0. Those sequences call vg_lite_hal_delay; the same tree’s drivers/gpu/vglite/vg_lite_hal.c implements it with msleep. This source behavior explains why repeatedly rebuilding the context can add waits unrelated to the actual draw. It is not a claim that every measured nanosecond belongs to one kernel sleep.

The next experiment should retain one serialized context across completed passes, close it after the final renderer owner is destroyed, and preserve failed-finish quarantine. It must repeat the actual pixel and recovery checks. The current path spends more process CPU as well as more wall time than its paired Pixman fallback. No offload benefit, ordinary-shell comparison, optical latency, IRQ accounting or performance acceptance is claimed from this pair.