[TibixDev]

What

Maintain draw-register fingerprints on writes, using the current draw dependency ranges. Runtime user words stay live; device serial and shader registry identity remain in the final key. Copy, restore, erase, clear and PM4 writes keep the fingerprint consistent. APS5_VERIFY_REGISTER_KEYS=1 checks it against full recomputation.

Rebuilt directly on main d146ec1b6, removing the old #1311 stack and its draw-plan API. This is a single-topic, seven-file change using the current prepared-shader draw path.

Tested

Fedora 44, GCC 16.2.1, Release:

• 30,000 differential mutations plus dependency ranges, masks, missing-versus-zero registers, retained write references, copies/moves, context restore and clear.
• Register-key, PM4, graphics and prepared-graphics CTests all pass (3.98 seconds total). Debug recomputation enabled; GPU cases use RX 6600 RADV with core/synchronization validation.
• Separate benchmark executables compiled with the exact main and proposed register containers, including write maintenance. Six runs per side in ABBA order, CPU 3 pinned, process CPU time, 32,768 lookups per case. With 24 live user words and 0/1/16/64/256 additional writes, main medians were 1.386/1.323/1.416/1.739/3.028 microseconds; proposed medians were 0.144/0.155/0.275/0.632/2.032 microseconds. These are warm component costs, not frame-time improvements.
• Conventions and diff checks pass.

No new isolated in-game performance claim. Plain main currently stops this Skyrim boot at unsupported NGS reverb channel conversion; a shared experimental NGS integration baseline is being checked separately. Windows and full platform CI were not run locally; those require the repository's collaborator-dispatched workflow.

Checklist

• Based on main; prerequisite and related work checked.
• One topic.
• No new code comments.
• Unsupported paths still throw; UI stubs are documented.
• No notes, logs, or images in the repository.
• No new third-party code.
• General behavior.
• Depends on: none.
• AI-assisted: yes.

Title artifact correction

A binary-hash audit found that the earlier title runtime copies all contained the same patched main AGC library. Those runs establish the main + experimental NGS baseline only. They do not validate #1312/#1313 in-game. The separately built CTests and CPU benchmarks used the new unpatched libraries/source and remain valid. The earlier PR-specific title claim is withdrawn; corrected results follow.

Verified runtime rerun

Corrected runtime result: the NID-patched #1312 library has SHA-256 a15e22c1b1305f79ab39089c7640ab9b6dd0fb9e6bffc0910d053cddfc7d51eb, differs from baseline, and contains the fingerprint verification code. /proc confirms the game loaded that file and the shared experimental NGS library. A fresh 120-second run with APS5_VERIFY_REGISTER_KEYS=1 reported at least 1,706 register-key lookups, no fingerprint mismatch, and presented frames. It stopped at the imposed timeout (124). No performance comparison is claimed with this verification enabled.

[TibixDev]

Rebased on c081eb90. Updated #1311 after #1134 merged. Non-indexed mesh draws retain the upstream null index descriptor; retained plans and their invalidation coverage remain intact.

Fedora 44 Release: agc_driver_draw_plan_tests, agc_driver_tests, agc_driver_graphics_tests, agc_driver_pm4_tests build; the targeted tests and contribution checks pass. Windows CI is pending.

[adriadam10]

Not mergeable as is, and I can't rebase it mechanically:

• It conflicts with main: #639 replaced compileDrawStage with materializeDrawStage (built from invocations prepared at registration), so the draw path this PR extends no longer exists. As @oneandonlydean already noted, the retained compile results from #1311 have to be redone on top of it, and this PR sits on that work.
• Please rebase the stack bottom-up (#1311 first, then this PR) onto current main and keep this PR to its own commit.
• Please post before/after numbers for this change alone against current main (the fingerprint maintenance on register writes), not against the pre-#639 tree: with shader compilation at registration, the per-draw cost profile changed.
• Re-run the draw-plan, driver, graphics and PM4 tests after the rebase, including the debug fingerprint-vs-recomputation check.

Happy to review again once it's on top of main.

[TibixDev]

Reworked this directly on main d146ec1b6 in 748d8d3af; #1311 and the old draw-plan stack are no longer dependencies. The fingerprint now feeds the existing draw key, preserving the current device/shader registry inputs and prepared-shader path.

Register-key differential/mutation tests, PM4, graphics and prepared graphics pass with APS5_VERIFY_REGISTER_KEYS=1; GPU tests use RADV core/synchronization validation. The new proxy also preserves compound writes and snapshot reads used by the graphics fixtures.

For an isolated comparison, I compiled separate benchmarks using main's exact register container and this branch's container. Six ABBA runs per side, pinned CPU, process CPU time, including writes: 24 live user words cost 1.386 → 0.144 µs; with 256 additional writes, 3.028 → 2.032 µs. Intermediate 1/16/64-write cases also improve. These are warm component numbers; the body no longer relies on combined-fork Helgen evidence or implies a measured title gain.

Full Linux/Windows CI still needs the collaborator-dispatched workflow. AI-assisted update.

Artifact correction: the title runtime copies were subsequently found to contain the same patched main AGC library. The saved PRESS ANY BUTTON frame therefore validates the main + experimental NGS integration, not #1312 or #1313. The CTests/CPU benchmarks used the new unpatched builds and remain valid. The earlier PR-specific title claim is withdrawn; corrected results follow.

Corrected runtime result: the NID-patched #1312 library has SHA-256 a15e22c1b1305f79ab39089c7640ab9b6dd0fb9e6bffc0910d053cddfc7d51eb, differs from baseline, and contains the fingerprint verification code. /proc confirms the game loaded that file and the shared experimental NGS library. A fresh 120-second run with APS5_VERIFY_REGISTER_KEYS=1 reported at least 1,706 register-key lookups, no fingerprint mismatch, and presented frames. It stopped at the imposed timeout (124). No performance comparison is claimed with this verification enabled.

[github-actions]

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.

Pull request	Conflicting files (lines on main)	Same exports	Same new files
#4855	core/libs/prx/libSceAgcDriver/tests/Pm4.cpp L4
#4943	core/libs/prx/libSceAgcDriver/Execution/src/Driver/Draw/DrawRegisterKey.cpp L8-9, L28-38

[adriadam10]

Checked on current main: merged tree builds, agc_driver_graphics, agc_driver_pm4, agc_driver_prepared_* and the new agc_driver_register_key pass. The new test does not compile without the change, and dropping the fingerprint update in Registers::erase makes it fail ("incremental fingerprint missed a mutation"). The Reference proxy also keeps any raw uint32_t& write from bypassing the fingerprint at compile time.

Note: #4943 touches DrawRegisterKey.cpp too and conflicts with this one, so whichever lands second needs a rebase. Merging once the full Build is green.

[adriadam10]

The full Build failed on Windows only in agc_driver_register_key (exit code 0xc0000135, DLL not found at start-up); the other 1051 tests passed. The new test target linked only libSceAgcDriver, while the sibling AGC driver tests (agc_driver_pm4_tests, agc_driver_graphics_tests) also link libc and libkernel.

Added on top of your commit, rebased on current main:

• core/libs/prx/libSceAgcDriver/CMakeLists.txt: agc_driver_register_key_tests now links libc and libkernel like its siblings.

Merging once the full Build is green on the new head.

[adriadam10]

The full Build on the previous head still failed on Windows only in agc_driver_register_key (0xc0000135, DLL not found at start-up). Linking libc and libkernel was not enough: the sibling AGC driver tests also get PATH=path_list_prepend:$<TARGET_FILE_DIR:libc> on Windows from the agcTest loop, and this test did not.

Added on top of your commit, rebased on current main:

• core/libs/prx/libSceAgcDriver/CMakeLists.txt: agc_driver_register_key prepends libc's directory to PATH on Windows, like its siblings.

Merging once the full Build is green on the new head.