Repository navigation
test(agc): agc_driver_pipelined_dispatch_item fails in about 1 of 70 Windows runs when VirtualAlloc puts its two targets in different 4 GiB regions #9680
Description
What happens
agc_driver_pipelined_dispatch_item sometimes stops before it submits anything: pipelined dispatch: the two targets do not share their high address dword.
Title, test or build step
agc_driver_pipelined_dispatch_item (agc_driver_pipelined_dispatch_tests.exe)
AnyPS5 commit
main d6eafa5
System
Windows 11 Pro 26300, GeForce RTX 3080 (NVIDIA 616.92), WinLibs GCC 15.2.0 r7, Release
Steps to reproduce
Run the test on its own many times, each run in its own ctest call (under a second each):
$env:ANYPS5_REQUIRE_VULKAN = "1"; $env:ANYPS5_NO_SHADER_CACHE = "1"
foreach ($i in 1..100) { ctest --test-dir build -R "^agc_driver_pipelined_dispatch_item$" --output-on-failure }Log
pipelined dispatch: the two targets do not share their high address dword
What was checked
- On
maind6eafa5 it failed 2 of 100 runs, and 1 of 40 in another batch, always with this message. It also failed once in a parallel^agc_run with fix(agc): round 16-bit float image stores toward zero #8998 applied (1 of 20 in isolation there); that PR touches neither this test nor its allocations. NewBlockreserves each 64 KiB block with its ownVirtualAlloc. The table rounds alternate betweenTargetAandTargetB, andTableFillsswitches the table's buffer descriptor by rewriting only its low dword, so the test needs both targets in the same 4 GiB region and checks that atPipelinedDispatch.cpp:176.- With a local
fprintfof the six block addresses (not committed), 3 of 279 runs failed. In each,TargetBlands in the next 4 GiB region after a gap, not next toTargetA:In the three passing runs I printed, the six blocks are consecutive, for exampleTargetA=000001bce8dd0000 TargetB=000001bd2b1e0000 TargetA=000001f5e7280000 TargetB=000001f63d040000 TargetA=0000024cea1f0000 TargetB=0000024d2c600000Input=000001d792500000toTargetB=000001d792550000. - No other test has this check (
git grep); it came with perf(agc): dispatch-cache recipe hits as ordered commit items #1747. - One way out: take both targets from a single
VirtualAllocof three blocks and use two neighbours that don't cross a 4 GiB boundary. - Not tried: Linux, where
NewBlockusesaligned_alloc.
AI-assisted: yes
Activity
MarcZX10 commentedon Oct 22, 2026
The proposed fix is in open PR #9696. It reserves one contiguous three-block target region, selects an adjacent pair that stays within one 4 GiB address window, and now includes Windows
libclinkage plus unwind configuration for the helper test target. The PR author reports Linux coverage; an independent Windows forced-boundary reproduction passed after the change, but the helper target’s MinGW/DWARF link and unwind issue was found in review and is now addressed. I also ran the helper executable locally with GCC 12.2 (-Wall -Wextra -Werror); it passed. I could not run the Vulkan runtime test on this host. Please keep the issue open until the PR’s updated checks and Windows validation complete and the fix reachesmain.MarcZX10 commentedon Oct 22, 2026
I updated #9696 to add the Windows test
PATHmodification for$<TARGET_FILE_DIR:libc>, matching neighboring libc-linked AGC tests, and merged current upstreammaininto the branch. New head:f24743f75078980bee1b8b0cc4ac03c8d1903520. I reran the standalone address-window helper test on this head (target address window test passed), plus diff-check and conventions; the full Vulkan dispatch runtime test is not available in this host environment. The PR conventions and labels checks are queued. I attempted the repository’s documented manual build workflow, but GitHub returned HTTP 403 because this account lacks repository admin rights, so no hosted Windows run was dispatched. Please keep the issue open until the Windows CI/test path succeeds and the fix is merged.MarcZX10 commentedon Oct 23, 2026
The updated #9696 head
f24743f75078980bee1b8b0cc4ac03c8d1903520now has a reported forced-boundary red-to-green run on RTX 3070 and lavapipe, in addition to the focused address-window helper checks. It also adds the Windows libc test PATH setup identified in review. The focused hosted conventions/intake/labels/report checks pass; product jobs are skipped by repository policy. I have not run the Vulkan test locally. Keeping this issue open until the fix merges is still appropriate.MarcZX10 commentedon Oct 23, 2026
The PR now has independent Windows 11 / MinGW / RTX 3050 results on the updated
f24743f750head: the focused pipelined dispatch test passed 1,000/1,000 on both main and main+PR, and the new address-window helper builds and passes on Windows. This validates the runtime test path but did not reproduce the intermittent failure in the main control, so the forced-boundary Linux red/green remains the direct regression reproduction. #9696 is still open; keeping this issue open until the fix merges is appropriate.