Guard cached compute precompilation against driver faults #758

Closed

klimovich008 wants to merge 3 commits from codex/guard-compute-precompile into main

klimovich008 commented

2026-11-08 20:10:34 +00:00

(Migrated from github.com)

What this fixes

On Windows, a driver access violation (0xC0000005) during cached compute precompilation terminates startup. This call bypasses the exception guard already used during shader preparation and runtime compilation.

Why merge

One-line change. Reuse the existing guard; non-Windows behavior stays the same.
Preserve the cache, creation arguments, successful publication and failure cleanup.
Keep startup limits, defaults and environment overrides unchanged.

Change

gpu/vk/vk_draw.cpp: use guarded_compute(pcache, &ci, &made).
tests/test_precompile_compute_guard.py: test the actual wrapper and worker branch with injected Windows faults.
tests/run_all.py: run the check on Windows and record a Linux skip.

Tested

On Windows: python -I -B tests/run_all.py --strict precompile_compute_guard; checks success, access violations, ordinary errors and abandoned-pipeline cleanup.
Fails on the old code, passes now. All five cases pass after the change.
Local Windows/Ryzen 7 5700X3D: the injected access violation kills the old child; the fixed child exits normally. No GPU driver is invoked.

All seven CI checks pass after syncing with current main, including the Windows regression and both platform builds.

Not covered

Real GPU startup validation is pending. This does not identify the cause of the reported RX 9070 XT startup abort.
Driver hangs, fast-fail termination and production worker scheduling are outside this test.

No game data is included.

## What this fixes On Windows, a driver access violation (`0xC0000005`) during cached compute precompilation terminates startup. This call bypasses the exception guard already used during shader preparation and runtime compilation. ## Why merge - One-line change. Reuse the existing guard; non-Windows behavior stays the same. - Preserve the cache, creation arguments, successful publication and failure cleanup. - Keep startup limits, defaults and environment overrides unchanged. ## Change - `gpu/vk/vk_draw.cpp`: use `guarded_compute(pcache, &ci, &made)`. - `tests/test_precompile_compute_guard.py`: test the actual wrapper and worker branch with injected Windows faults. - `tests/run_all.py`: run the check on Windows and record a Linux skip. ## Tested - On Windows: `python -I -B tests/run_all.py --strict precompile_compute_guard`; checks success, access violations, ordinary errors and abandoned-pipeline cleanup. - Fails on the old code, passes now. All five cases pass after the change. - Local Windows/Ryzen 7 5700X3D: the injected access violation kills the old child; the fixed child exits normally. No GPU driver is invoked. All seven CI checks pass after syncing with current main, including the Windows regression and both platform builds. ## Not covered - **Real GPU startup validation is pending.** This does not identify the cause of the reported RX 9070 XT startup abort. - Driver hangs, fast-fail termination and production worker scheduling are outside this test. No game data is included.

NotRexed commented

2026-11-09 05:17:01 +00:00

(Migrated from github.com)