Repository navigation
Conversation
|
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.
|
4cc4931 to
3bac469
Compare
|
Temporarily closing this draft to free one of my 20 open-PR slots for a reproduced cache-header arithmetic correction following #6296. This first-use pipeline work remains on fix/lazy-pipeline-abi at 3bac469, and its prior testing integrations are preserved. It still needs an isolated main-versus-change title performance comparison before becoming ready. Reopen when a slot is available; no branch deletion. |
github-actions
Bot
added
perf
3bac469 to
d5ec29e
Compare
d5ec29e to
22cee06
Compare
22cee06 to
c8b4a44
Compare
|
Rebased without merge commits to Both main and change pass the same eight focused Vulkan checks, including the upstream deferred compute and concurrent lazy cases, with validation verified loaded and zero VUIDs. The added first-draw lazy pipeline fixture passes too. Full Release libraries build; conventions and whitespace pass. Pinned receipts and logs. The existing technical draft remains because the required matched-title performance evidence is still outstanding. These focused checks do not establish a title loading or FPS gain. |
13843b1 to
5fe8ec8
Compare
|
Current-main conflict refresh, 2026-10-27 10:18 UTC pass: exact head Linux Release, RX 6650 XT/RADV Mesa 26.2.4, matched SPIRV-Tools-disabled builds. Matched main passes 14 groups × three repetitions (42 passes). Candidate passes 15 × three (45), and all 45 pass again after restoration. With only main's production ShaderRegistry.cpp and candidate tests, the 15 groups run once: 14 pass and only lazy-pipeline-ABI fails at the expected eager-preparation assertion. Restored source matches the candidate and the AGC driver is byte-identical. Actual Khronos insertions: 84 main, 90 initial candidate, 30 negative, 90 restored. Every physical fixture has actual RX 6650 XT selection and Khronos validation loading verified; CPU/mock checks have no GPU-execution claim. Required Vulkan and shader caches disabled; zero skips, VUIDs or missing-layer failures in the accepted runs. Conventions and whitespace pass. All complete test harnesses and cleanup ran through the shared FIFO in service-owned PTYs and finished normally. No new title, console, Windows, visual, readback, FPS or performance result is claimed. Older observations below retain their original exact pins. Exact patches/pins, integration and source proofs, builds, verbose matched tests, validation review, hashes and terminal service/FIFO receipts. SHA-256: AI-assisted: yes. |
5fe8ec8 to
b3ce8e1
Compare
|
Refreshed on main
Loader logs verify validation insertion (84 main, 90 initial candidate, 30 control, 90 restored), the selected physical GPU, zero VUIDs, missing-layer diagnostics or skips. The complete harnesses, including source control/restoration, ran through the shared FIFO under service-owned PTYs. Conventions and whitespace pass. This remains a technical draft because a matched-title loading/performance benefit is still unqualified; no new title, Windows or speed result is claimed. Exact source/dependency pins, source proof, builds, service/FIFO receipts, controls, logs and JUnit, SHA-256 AI-assisted: yes. |
github-actions
Bot
added
needs rebase
b3ce8e1 to
cf0094f
Compare
|
Refreshed head Both production early-lazy guards are retained. Main’s new budget-only fixture never requests an ABI, so the transplanted ResolveShaderAbi setup is now conditional on Fresh matched Linux Release builds on AMD Radeon RX6650XT / RADV, with SPIRV-Tools enabled: main 48 passes and change 51 passes (three repetitions of each focused group), zero skips and zero VUIDs. Actual Khronos validation-layer instance insertions are 96 on main and 102 on the change. CPU-only groups are not counted as GPU coverage. Conventions, whitespace and submodule-pointer checks pass. Later-main merge previews were clean; those preview trees were not executed. Every GPU harness used the shared reservation and launch queue. Evidence: raw logs, JUnit and source receipts; SHA-256 AI-assisted: yes (Codex). |
github-actions
Bot
added
needs rebase
(cherry picked from commit 22cee06)
cf0094f to
496b51b
Compare
|
Refreshed on main |
github-actions
Bot
added
perf numbers missing
|
This pull request is labeled |
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Cold shader cache A/B (Linux), at head Setup:
That is a hint that deferring graphics-stage ABI preparation eases the worst cold-cache stalls. Five pairs and 2 s samples don't establish it. The PR stays a draft until a larger or trace-based measurement does. AI-assisted: yes |
What
Draft: matched-title performance benefit remains unproven.
With
APS5_LAZY_SHADER_PREPARE=1, defer artifact preparation in the graphics-stage and rectangle pipeline ABI resolvers until first use, after registered-header and device validation. Default eager behavior is unchanged. Main's existing single-stage ABI planning, saved deferred ABI state and synchronization are preserved; the production change is two early-return guards.Refreshed as two linear commits on main
cf79276fee2f042feb6a7c71d0d6a27e9cfe2592; head496b51b41ad5161568b089e7e75de281119b033c. The new ABI regression now uses main's per-test CMake fragment layout. No private timing probe or title integration is part of this branch.Tested
Linux Release/GCC 16.2.1, AMD RX 6650 XT/RADV Mesa 26.2.6, Linux 7.2.11. Both full SDKs/relinkers and focused executables built, with matching dependency gitlinks and patched/unpatched ELF build IDs. Matched focused checks ran through the shared FIFO with
ANYPS5_REQUIRE_VULKAN=1,ANYPS5_NO_SHADER_CACHE=1and actual Khronos validation loading verified:Exact source patch, dependency pins, builds, verbose logs, JUnit and terminal receipts, SHA-256
7748e20d19a542d2b31fb8e4fb097a95118807d8a1389e30faf4cc1f2e1f5d3f(uploaded/downloaded and internal hashes verified).No new Windows, NVIDIA or title speedup is claimed by these checks. The Smurfs – Dreams PPSA21607 01.002.000 cold-start pair is complete. Control
62bbc741fb2fis maincf79276fee2fplus the same #6959/#13934/#6782/#6913 prerequisites and a private common timing probe; changea7cae4a3bed6adds this PR. Normal flip scheduling, lazy flag enabled in both arms, fresh private AnyPS5/Mesa shader caches disabled, identical initial writable hashes, same scheduled input, 3840×2160. Both exit0 in about98s, restore kernel settings and leave original inputs/SDK hashes intact.No reliable title loading or gameplay FPS benefit is established by this single pair. Inclusive phase sums overlap and include concurrent work; deferred ABI time is not all removed. Only1.29ms of the control's graphics-stage ABI work precedes its first Present, so the136.58ms process-start difference cannot be attributed to that work. Both reviewed90s captures show save selection; the80s captures differ, and inactive-window capture freshness could not be verified. Mixed startup scene histories make the0:60 trace frame distributions unsuitable for a steady-scene FPS claim. The private measurement source is identical between arms and is not in this PR.
Both official traces/summaries, reviewed captures, source-isolation proof, exact private probe patch, reports and timing counters, SHA-256
35a5f2e433624fd81f590e882ac9743fb55745fd161252bae302d881489f2fac(download/internal hashes verified).The historical Resident Evil Village cold-start pair, main
a6dcea9279b5versus22cee06b0105, did not establish a loading gain. Its startup scene histories differ, so its frame distributions do not support a gameplay speed comparison.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)