Repository navigation
test(agc): bound the storage cache budget regression workload - #2027
Conversation
|
I reviewed the PR metadata and the full diff. The PR currently includes substantial implementation changes from #1971, #1265, and #1986, all still open, alongside the bounded-test work. Could you isolate this test change on top of the prerequisite stack (or rebase it after those prerequisites merge) so this PR’s diff contains only the test topic? I’ll hold off on approval until the diff is isolated and Linux/Windows finish for that head. |
881e519 to
68072d6
Compare
|
Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.
|
|
Native Windows run, since no Windows results were posted. Setup: Windows 10 (19045), RTX 3060 Laptop (NVIDIA 616.92), MinGW-w64 GCC 15.2.0 (WinLibs r7), Release. Head
No title was run. 🤖 Addressed by Claude Code |
What
agc_driver_storage_cache_budgetuploads and flushes 64 MiB images to test a multi-gigabyte cache. Cold Linux lavapipe runs in #1059/#1971 took 40–60 seconds, exceeding the 30-second test limit, and CPU devices skipped the eviction assertions.Use the existing
APS5_TEXTURE_CACHE_MIBoverride for a 32 MiB test cache and 512×512 R32_UINT images (1 MiB each). Size the retention and eviction sets from the actual accounted allocation. Keep shader marker readback, object-identity, retained-entry, exact two-oldest-entry eviction and cleanup checks; execute eviction on CPU Vulkan too. Set the test's CTest timeout to 30 seconds. Existing graphics unit tests retain coverage of the physical-heap budget formula.#1059 has merged. This branch has been rebuilt directly on
maind70b8998: one commit, two files. The final diff contains only the bounded storage test and its timeout. All #1971, #1265, #1986, #1999 and #1913 prerequisite changes have been removed; none is needed for this patch. This addresses the review request to isolate the diff.Duplicate review inspected #1059/#1971 and #1971's
e3180e13update: that update fixes sampled-texture identity and does not bound this storage test. The broader #2001 removal proposal is not an implementation dependency.Tested
Exact source head
68072d60: independent validation passes on native GitHub Linux and Windows runners, Release / GCC, including full default builds, all-library packaging and three repetitions of the relevant runtime tests under a 30-second limit. Linux lavapipe executes the storage test three times with distinct empty Mesa caches: 4.63, 1.43 and 1.37 s; graphics unit tests also pass. Hosted Windows runs the graphics tests and skips the storage Vulkan test because no usable device is available. No Windows GPU execution is claimed.These are fork validation jobs, not collaborator-dispatched upstream checks. Unrelated tests from the full suite were not run in this validation.
Earlier pre-PR Linux validation on the prerequisite integration branch took 0.44, 0.42 and 0.40 s for separate cold caches. Ignoring the configured cache budget deliberately failed the eviction assertion in 0.42 s, confirming that the test checks eviction. Earlier validation. This historical result is separate from validation of the newly isolated head.
Conventions and
git diff --checkpass. No production cache policy or shader semantics change. Windows hosted Vulkan skips, if any, are disclosed in the current validation result; no physical-GPU or console testing is claimed.Checklist
mainsnapshotd70b8998; no other open PR implements this change