Repository navigation
fix(agc): remember the differing statistics' value sets by hash - #8263
Merged
oneandonlydean merged 1 commit intoOct 21, 2026
Merged
oneandonlydean merged 1 commit into
oneandonlydean merged 1 commit into
Conversation
With APS5_PROFILE_DRAW, every dispatch that misses because its captured SRT words differ copies the entry's runs and words into a ring of the last three value sets per dispatch key, for up to 16384 keys. The ring only answers whether a new value set equals one of them, so it now keeps a 64-bit hash of each set instead of a copy. Ghost of Tsushima captures SRT words of more than 1 MiB per compute dispatch, and a profiled run held 352 MiB of these copies 274 s in. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
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.
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.
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
With
APS5_PROFILE_DRAW,Driver::classifyDifferingsorts each dispatch-cache miss whose captured SRT words differ from the entry's. To count how often a miss brings back one of the last three value sets of its key, it copied the entry'srunsandwordsinto a ring of three per dispatch key, for up to 16384 keys (priorValueSetsis cleared only when it holds more keys than the dispatch cache has entries). The ring is only compared for equality, so it now keeps a 64-bit FNV-1a hash of each value set (the runs and the words, each with its size) instead of a copy. The counter reports the same numbers. A hash collision could add a match with a chance of about one in 2^64.Why
Ghost of Tsushima (PPSA02225) captures more than 1 MiB of SRT words per compute dispatch. I took a profiled run at 16 GiB on
main-based libraries (97f525ba19 plus #8213 and #6982;DispatchCacheStatistics.cppandpriorValueSetsare the same as onmain). Joining the live allocations 274 s in with their allocation stacks showed 352 MiB in these copies: 134 + 90 MiB from the vector copies inclassifyDiffering, and 77 + 51 MiB in thestd::dequeinsertion it calls. Profiled runs are how titles are measured for pull requests, so the copies took memory from the run being measured. In this title they used memory that a run without profiling has.Tested
Ubuntu 24.04 (Linux 6.8), GCC 13.3, NVIDIA RTX 4070 Ti, driver 595.99.02.
agc_driver_dispatch_differing_value_sets. It uses one dispatch key, a kernel that loads its output V# through the SRT pointer, and an SRT that alternates between two value sets (the V#'s record count, 64 and 65).APS5_DISPATCH_VARIANTS=1makes every dispatch after the first a differing miss. The test checks that the four misses are classified and that the three bringing back a value set seen before match it. It reads the counters through the newDriver::DispatchCacheCounters(), the dispatch counterpart ofDrawCacheCounters(). The counters are reset by the 10 s profile report, so the test sums deltas across a reset.mainwith only the accessor and the test: 4 classified, 3 matched, 3 of 3 runs. This PR: the same, 3 of 3. So the statistic is unchanged.valueSet(old) + 1), it fails with 0 matched.VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation ANYPS5_NO_SHADER_CACHE=1 ANYPS5_REQUIRE_VULKAN=1 ctest -R '^agc_driver_(dispatch|image_table|compute|address_passes|draw_cache)' -V.main1edbe4b passed 20 of 20, and this change at 8d99dad5ba (the same diff before the rebase onto 1edbe4b) passed 21 of 21, the new test included. Neither log reports the layer missing, and neither has aVUID-error. After the rebase the new test passes 3 of 3.tools/check_conventions.py --base upstream/mainpasses.tools/pr_overlap.pyonly lists open PRs it couldn't compare, onDriver.hppandDispatchCacheStatistics.cpp. Of those, feat(agc): look up table images at run time through a device image table on the GPU heap (APS5_GPU_PATH) #7642 is the one that edits the statistics file, and it only changes the report line.Not tested: a profiled title run with this change. On this machine, Ghost of Tsushima at 16 GiB now stops at the host memory guard (126 s and 151 s today) before the point measured above.
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)🤖 Generated with Claude Code