mirror of
https://github.com/Droid-VM/crosvm.git
synced 2026-11-17 06:58:03 +00:00
B8 §6 set AWB_MODE, EFFECT_MODE and SCENE_MODE through V4L2 while a capture was running and the backend reported all three "dropped", with the capture result carrying the neutral value -- while the same three set before STREAMON were "echoed". The entries were never lost: the instrument was reading too early. A repeating request replaced mid-stream does not reach the sensor at once, and the results still in the pipeline legitimately carry the old settings; B8's own camera_probe run shows the new value arriving five to seven results after the submission (16_d36_probe.txt: "result 19: AWB_MODE Some(Fluorescent)" for a set made twelve results earlier), and ECHO_AFTER was six. The verdict was read one result before the answer, three times out of three. So stop counting. `on_capture_completed` is handed the request the capture was taken with, and `ACaptureRequest_getConstEntry` reads it: a result whose own request does not carry the value is simply an older capture and decides nothing, and the first result whose request does carry it is the camera's answer to that entry. `CaptureResult::requested_awb_mode` and its three siblings expose that, and ModeEcho waits for it rather than for a frame count -- with ECHO_DEADLINE (30 results, a second) left as the bound past which no request ever carried the value, which would be an entry that never reached the camera and so a bug of ours. A result that does not echo what the camera ran is now a warn!, once per control per stream (design §7.1's D36 advisory: the guest reads back what it set while the camera uses something else, which is worth saying out loud), and a log line thereafter. camera_probe's `--watch` measures the same thing without a guest -- every set is mid-stream, and each line says how many results the request took to come back -- so B9 can verify D47 from the host. Two more from the same B8 section. `Geometry::area_of` still logged nothing, so "the host log shows active-array coordinates" could not be performed for D34 (D39): every *_REGIONS set now prints one info! with the guest rectangles in REGION_SCALEths of the frame, the converted active-array rectangles, the array itself and the zoom ratio in force. And a white balance or an exposure set while a scene mode is in force is warned about once, because Camera2 disables the individual 3A controls under CONTROL_MODE USE_SCENE_MODE (NdkCameraMetadataTags.h, ACAMERA_CONTROL_MODE and ACAMERA_CONTROL_SCENE_MODE). Verified with deploy/vpu/harness.sh acb (a clean type-check of both files and the probe), vmt 119/119 unchanged, rustfmt --check adding no hunk, and an aarch64 soong build of camera_probe (logs/probes_f11.log, exit=0). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011BkN7vYhuc1w9iVuqsHzSn |
History
|
|
|---|---|---|
| .. | ||
| src | ||
| Android.bp | ||
| Cargo.toml | ||