Repository navigation
Conversation
|
Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.
|
jotitadev
left a comment
jotitadev
left a comment
There was a problem hiding this comment.
Apple GPU run of the new tests, with and without #6363.
Machine: Apple M2 Max, 32 GB, macOS 27.2, x86-64 build under Rosetta 2.
Method: macos-port @ b7eb6b6 + ImportWatched + this PR's diff (head 8e3755364), not main. The port's test Device has no context.queueFamily, so Timestamps() reads the family index the test Device already picks. No other change to the test. #6363 applied at 44d31ef9d for the second run.
Drivers: MoltenVK 1.4.2 (Vulkan SDK 1.4.363.0), and MoltenVK main + SPIRV-Cross #2704. Driver checked with VK_LOADER_DEBUG=driver.
MoltenVK reports timestampValidBits = 64 on the queue, so neither case skips. Both cases run their 3 rounds.
| Source | MoltenVK 1.4.2 | MoltenVK main |
|---|---|---|
| port + these tests | complete passes; incomplete passes | both pass |
| port + these tests + #6363 | both pass | both pass |
Every run finishes in under 2 s; the 15 s timeout never comes close.
To see why there is no hang, I temporarily logged the vkGetQueryPoolResults result in Recorder::readGpuTiming. Without #6363, so still with VK_QUERY_RESULT_WAIT_BIT, MoltenVK returns VK_NOT_READY for the incomplete round in about 0.01 ms instead of blocking on the unwritten query. The complete rounds return VK_SUCCESS. With #6363 the results are the same: complete rounds still return VK_SUCCESS and keep their timings without the wait bit.
So Apple behaves like lavapipe here: the incomplete case does not fail without #6363, and the negative control only reproduces on RADV. That matches the PR description; this adds a third driver to it.
GPUs tested: Apple M2 Max, MoltenVK 1.4.2 and MoltenVK main (+ SPIRV-Cross #2704), macOS 27.2.
Written with AI assistance (Claude Code). I ran the builds and tests above on my machine.
|
Temporarily closing this dependent draft to stay within the 20-open-PR author limit while submitting a verified Linux SDK allocator ownership fix for Silksong (PPSA12544, discussion #1519). This test remains useful and still depends on #6363; closing it does not mean the test or issue is resolved. The branch |
8e37553 to
e5e603d
Compare
This branch has not been deployed
2 participants
What
Add real-queue regression coverage for the timestamp-query hang fixed by #6363. The new incomplete-range case opens two timing ranges, deliberately omits the first range's end timestamp, completes the second and submits the batch. Batch completion callbacks must run and the completed-batch record must remain available. Complete batches before and after the incomplete batch must still report valid GPU timings. A separate all-complete case provides a passing control.
This PR adds tests only. The runtime fix #6363 merged into main on 2026-10-18; its implementation and authorship remain in that PR. Reproduction source including that fix is available on the validation branch.
The test checks the selected queue's timestamp support and returns skip 77 if unavailable. CTest enables GPU profiling and bounds either case to 15 seconds. Vulkan specifies that querying unavailable results without WAIT may return VK_NOT_READY: vkGetQueryPoolResults.
Tested
Arch Linux, Ryzen 9 7950X, Release with bundled dependencies, historical baseline main
b7cedd900d67a94453b8a45eec9b5623532a0441.44d31ef9d)Both Vulkan drivers are Mesa 26.2.4-arch1.1; lavapipe uses LLVM 23.1.1. The test executable is byte-identical between the final negative control and fixed runs (SHA-256
6d851fd74156f7f36c75f59ec4111ce6e2d906697a0d6c5adfd34035c7bcbab5); only the runtime library changes. This catches a RADV failure that the software driver does not reproduce.Independent Apple GPU results from jotitadev: on an Apple M2 Max, macOS 27.2, x86-64 under Rosetta 2, both cases passed with and without #6363 using MoltenVK 1.4.2 and MoltenVK main plus SPIRV-Cross #2704. This was their
macos-portbuild atb7eb6b6bplus ImportWatched and this test diff (8e3755364), with a queue-family accessor adjustment; it was not a run of upstream main. They reportVK_NOT_READYreturning immediately for incomplete queries even with WAIT. These collaborator results extend the driver comparison but do not provide a negative control on Apple or replace the RADV reproduction. I did not run this Apple configuration.Full
libsandagc_driver_recorder_testsbuild passed. On the fixed validation source, four tests pass with zero skips on each driver:agc_gpu_timing_complete,agc_gpu_timing_incomplete,agc_unchanged_cpu_stamps,agc_recorder_read_list. Conventions, test registration and whitespace checks pass. No native-title or Windows run and no performance claim.Refresh after #6363 merged
Rebased the single test commit onto main
659230f7b676e8c848d70640f823342be6704526ase5e603d161f4ab1a0fcd81f0647c4cfd782a9a56. Fulllibsandagc_driver_recorder_testsbuild passed. UnderVK_LAYER_KHRONOS_validation, five checks pass with zero skips on RADV (0.16 s total) and lavapipe (0.20 s total): both timing cases, unchanged CPU stamps, recorder read list and recorder device selection. The historical negative control above remains separate from this current-main positive run. No native-title or performance claim.The overlap tool reports no comparable collisions. It cannot compare 53 already-conflicting PRs sharing the registration/recorder files; their changed lines were separately scanned and none changes the timing APIs or these timestamp cases.
Checklist
main; no other open PR implements the same functionsNotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file.mdfiles or images added to the repository (attach them to this PR)