Repository navigation
perf(agc): prebuild registration libraries on half the hardware threads by default - #10388
Merged
oneandonlydean merged 1 commit intoOct 27, 2026
Merged
Conversation
…ds by default Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
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.
|
Contributor
Author
|
Traces: main (e13027a): This PR (2a3cc34): |
3 tasks done
12 tasks done
Contributor
|
Windows run of this head
AI-assisted: yes (Claude Code) |
This was referenced Oct 28, 2026
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
On a device with graphics pipeline libraries, the registration-form vertex and pixel libraries are linked on the library prebuild pool. Since 361c7e2 that pool has one thread unless
APS5_PREBUILD_THREADSsays otherwise. On NVIDIA it is the startup bottleneck. With a warm shader cache, High on Life (PPSA15388) spends about 50 s with a single busy thread before its first present.perf record -gon that thread showsVulkanDevice::PrebuildGraphicsLibraries→Graphics::Pipeline::Prebuild→LinkPipelineFromLibraries→ the NVIDIA compiler, with 96% of its samples in the compiler.Change: the prebuild pool now defaults to the size of the registration preparation pool: half the hardware threads, at least two.
APS5_PREBUILD_THREADSstill sets the count, andAPS5_PREBUILD_THREADS=1gives the old behaviour.RegistrationPrepareStackcounts the new default, and thedocs/dev/ENVIRONMENT.mdrow describes it.agc_driver_shutdown_threads_queued_prebuildsnow runs withAPS5_PREBUILD_THREADS=1. It blocks one prebuild and checks that close drops the tasks queued behind it, and with more workers those tasks start right away on the other threads.Tested
Hosts:
Method: warm shader cache, no input. Time is measured from launch to the title's first present.
maine13027amain74f6fa8 / the same change on itmainc188e01APS5_PREBUILD_THREADS=16mainc188e01APS5_PREBUILD_THREADS=16APS5_PREBUILD_THREADS=1/=12main, one with this change) crashed after their first present, in a guest-side race of this title that predates both builds. Their first-present times are included.pipelineLibrary=1), but the start time doesn't change. Linking isn't the bottleneck there, so the change neither helps nor hurts.APS5_TRACE_SUMMARY_EVERY=10. Over the walking intervals:No regression.
Summary lines: PPSA15388 title screen,
APS5_TRACE_SECONDS=10:10, same machine. These are the runs whose traces are attached. Once the title is up, frame times and GPU time are unchanged:AGC tests:
maine13027a, the 39 tests matchingprebuild|registration_prepare|prepared_|shutdown_threadspass on both, with the same validation errors:vkCreateDevice-ppEnabledExtensionNames-01387×108 andVkPipelineColorBlendStateCreateInfo-pAttachments-00605×14.agc_drivertests were run withctest -V -j8onmain74f6fa8 and this change, withVK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation ANYPS5_NO_SHADER_CACHE=1 ANYPS5_REQUIRE_VULKAN=1, on the RTX 4090.VUID-:vkCreateDevice-ppEnabledExtensionNames-01387×1506,VkWriteDescriptorSet-descriptorType-00337×18,VkPipelineColorBlendStateCreateInfo-pAttachments-00605×18,VkImageMemoryBarrier-image-03320×6.agc_driver_sample_order,_grid_referenceand_recipe_depth_locationsfail on both. test(agc): read the sample order tests' depth through a sampled copy #9056 and fix(agc): bind the sample order depth set after its pipeline #9503 are open for them.agc_driver_pixel_lds, with a device fault. It fails the same way onmain74f6fa8 when run alone (3 of 3 runs, and 5 of 5 on this PR withAPS5_PREBUILD_THREADS=1), so it isn't caused by this change. It passed once inmain's parallel run.agc_driver_shutdown_threads_queued_prebuildsfailed with the new default ("closing ran the queued library prebuilds"). With one blocked worker, the other workers ran the queued prebuilds. Its environment now pinsAPS5_PREBUILD_THREADS=1. All nineagc_driver_shutdown_threads*tests pass.ShaderRegistry.cppfrommain,agc_driver_registration_prepare_stackfails with "the preparation and prebuild pools count 17 live threads of the 32 they started". With this PR it passes.tools/check_conventions.pypasses.tools/env_reference.py --checkreports onlyAPS5_NO_WGI, which is undocumented onmaintoo.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)agc_drivertests and titles); AMD Radeon Pro V620, RADV Mesa 25.2.8, Ubuntu 24.04 (title start only)🤖 Generated with Claude Code
Tested and written by Claude (AI assistant) on a contributor's machine.