Preserve partial overlaps with CPU-written report words #442

Merged

klimovich008 merged 1 commit from codex/protect-partial-report-writeback into main

2026-11-08 16:49:22 +00:00

klimovich008 commented

2026-11-08 02:01:51 +00:00

(Migrated from github.com)

What this fixes

Mirrored GPU writes can overwrite a newer CPU-written timestamp or query result when the copied range starts halfway through its 8-byte report word.

Why merge

One-line change. Start the ownership lookup at the containing report word so partial overlaps are excluded too.
Preserve every nonowned byte, including prefixes and suffixes around protected words.
Keep the existing write-back path, defaults and RX_VK_WB_REPORTS override unchanged.

Change

gpu/vk/vk_memory.cpp: use cpu_words.lower_bound((p.lo + at) & ~7ull) when cutting protected ranges.
tests/test_report_words.py: check the actual helper against expected bytes for partial starts, partial ends, adjacent words and all 4-byte-aligned ranges in the fixture.

Tested

python -I -B tests/test_report_words.py --cxx clang++ checks protected and nonowned bytes. Fails on the old code, passes now.
Matching Windows host and renderer build successfully.
RX 9070 XT, Windows: 182-second headless Kokiri smoke test, zero VUIDs. No game data is included.

Not covered

Tested on AMD only. The game run is a smoke check; the CPU regression reproduces the partial-overlap failure.
A partial overlap was not captured in the game run; no visual flicker fix is claimed.

## What this fixes Mirrored GPU writes can overwrite a newer CPU-written timestamp or query result when the copied range starts halfway through its 8-byte report word. ## Why merge - One-line change. Start the ownership lookup at the containing report word so partial overlaps are excluded too. - Preserve every nonowned byte, including prefixes and suffixes around protected words. - Keep the existing write-back path, defaults and `RX_VK_WB_REPORTS` override unchanged. ## Change `gpu/vk/vk_memory.cpp`: use `cpu_words.lower_bound((p.lo + at) & ~7ull)` when cutting protected ranges. `tests/test_report_words.py`: check the actual helper against expected bytes for partial starts, partial ends, adjacent words and all 4-byte-aligned ranges in the fixture. ## Tested - `python -I -B tests/test_report_words.py --cxx clang++` checks protected and nonowned bytes. Fails on the old code, passes now. - Matching Windows host and renderer build successfully. - RX 9070 XT, Windows: 182-second headless Kokiri smoke test, zero VUIDs. No game data is included. ## Not covered - Tested on AMD only. The game run is a smoke check; the CPU regression reproduces the partial-overlap failure. - A partial overlap was not captured in the game run; no visual flicker fix is claimed.

Commenting is not possible because the repository is archived.

No reviewers