[MarcZX10]

What

Adds per-group attribution for draw register-key cache misses: stable registers, user-data registers, auxiliary registers, device, and shader set, plus misses where none of those components changed.

This extends the existing APS5_TRACE_DRAW_CACHE report, which already distinguishes new-register misses, user-word/pointer-only misses, evictions, and keys seen but never inserted. It reuses traceDrawCache() and its 10-second report; there is no separate environment variable or stderr logging path. Shader-set identity is included so distinct snapshots are tracked separately.

Tested

• Linux: built agc_driver_register_key_tests and libSceAgcDriver.prx successfully with Ninja Release.
• APS5_VERIFY_REGISTER_KEYS=1 ctest --test-dir /tmp/anyps5-open-diagnostics-main-build -R '^agc_driver_register_key$' --output-on-failure — the test executable cannot enter its test body on this host: Xeon E5-1620 v2 lacks LZCNT, BMI1, BMI2, MOVBE, AVX2, and FMA.
• python3 tools/check_conventions.py --base upstream/main — passed.
• git diff --check upstream/main...HEAD — passed.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical-debt comments
• Unimplemented paths throw or are documented; not applicable
• No notes, investigation files or images added
• No third-party code added
• Generic behaviour, not title-specific
• Depends on: none
• Unblocks: none
• GPUs tested: n/a (diagnostic only)
• AI-assisted: yes

[fma-rc]

I ran this head merged into main 0aa4cb1 on Astro's Playroom (Linux, RTX 5080) with APS5_TRACE_DRAW_KEY_MISSES=1. About 145,000 draw-key lookups miss per 10 s there, and the whole run printed four lines:

[draw-key-miss] shader-set=cbf29ce484222325 changed=first-seen
[draw-key-miss] shader-set=cbf29ce484222325 changed=stable-registers,user-data-registers,device
[draw-key-miss] shader-set=cbf29ce484222325 changed=stable-registers,user-data-registers,device
[draw-key-miss] shader-set=cbf29ce484222325 changed=stable-registers

cbf29ce484222325 is the FNV offset basis. In drawRegisterKey, (*components)[4] is set to it and the shader identities are mixed into (*components)[3] only, so every draw has the same shader-set id and the four-lines-per-set cap stops the trace after four misses. It also means device reports a change whenever the shaders differ. agc_driver_register_key passes here.

[oneandonlydean]

mixShaderIdentity mixes into (*components)[3], the device slot, so (*components)[4] stays at the FNV basis for every draw (DrawRegisterKey.cpp:86-95). Every draw then has the same shader-set id, device changes whenever the shaders do, and the four-line cap ends the trace after four misses, as fma-rc's run shows. Please mix the shader identity into [4] and add a check in tests/RegisterKey.cpp that two shader sets get different ids.

[MarcZX10]

Fixed the component mix-up: shader snapshot identities and offsets now feed the shader-set key in components[4], while components[3] remains the device serial. Added a regression comparing two registries with different shader snapshots. Linux Release build of agc_driver_register_key_tests succeeded; focused CTest passed 1/1, and conventions passed.

[adriadam10]

Checked the new head: components[4] now carries the shader-set key and components[3] the device serial. agc_driver_register_key_tests builds and passes with APS5_VERIFY_REGISTER_KEYS=1, and the new shader-set check fails when the snapshot mix in DrawShaderSetKey is removed. Default behaviour is unchanged with the variable unset. Merging once the full Build is green.

[parl0v]

Ran this in a title: PPSA25344 (Sword of the Sea), title screen, main f0854887d + this PR at 2dd5ecdf4 (after the components[4] fix), APS5_TRACE_DRAW_KEY_MISSES=1, RTX 3060, Windows 10, 300 s. It builds and runs, and frame rate matched main (3.70 FPS both).

It printed 217 [draw-key-miss] lines from 57 shader sets (the per-set cap of 4 is hit by 52 of them):

changed=	lines
stable-registers,user-data-registers	81
user-data-registers	76
first-seen	57
no-component-change	1
truncated (changed=st, a line cut by interleaved output)	1

Among the sampled repeat misses, user-data registers changed in all but one, and stable registers changed along with them in about half. One note from using it: there's no per-10 s total, so the lines show which components change but not how many draws miss in all. A count line like #5408's [keymiss] summary would make it comparable across runs.

[adriadam10]

This now conflicts with main in DrawRegisterKey.hpp/.cpp, Driver.hpp and Draw.cpp, and it is not a mechanical rebase: #58b6ddf72 added APS5_TRACE_DRAW_CACHE (traceDrawCache(), absentUserWords / absentNewRegisters / absentEvicted counters) and changed RegisterStateKey/drawRegisterKey to return a shapeKey through the same parameter slot this PR uses for RegisterStateKeyParts / components.

Could you rebase onto current main and:

• reuse the existing traceDrawCache() gate and shape key instead of adding a second env var and a second miss path in Driver::draw, or say why a separate trace is needed;
• keep only what main does not already report (the per-group breakdown: stable / user-data / auxiliary / device / shader-set), and say in the PR body what it adds over the absentUserWords / absentNewRegisters split;
• rerun agc_driver_register_key_tests with APS5_VERIFY_REGISTER_KEYS=1 on the rebased head.

The components[4] fix and the new shader-set test looked right on the previous head; they only need to survive the rebase.

[MarcZX10]

Rebased onto current main and integrated the requested breakdown with traceDrawCache() and its existing 10-second report. The existing absentNewRegisters / absentUserWords / eviction accounting remains intact; the report now adds miss counts for stable registers, user-data registers, auxiliary registers, device, first-seen shader sets, and misses where no component changed. Shader-set identity remains in component slot 4, with a regression test for distinct snapshots.

Validation on the rebased head: the focused agc_driver_register_key_tests target builds successfully and conventions pass. CTest starts the binary but cannot enter the test body on this Xeon E5-1620 v2 because the host lacks LZCNT, BMI1, BMI2, MOVBE, AVX2, and FMA. I recorded that limitation in the PR description.

[fma-rc]

@MarcZX100 the force-push at 23:58 UTC left this branch with a single commit, fdba064 fix(libkernel): log resolved paths for failed opens, which only touches core/libs/prx/libkernel/File/src/Open.cpp and core/libs/tests/GuestFilesystem.cpp. The draw-key work and the traceDrawCache() breakdown you describe aren't in the diff any more (no traceDrawCache or register_key lines), so it looks like another branch went to marc/trace-draw-key-misses. Once the right commit is back I'll build it and run it on Astro's Playroom, where most draw-cache misses carry the same key as the set's last one.

[MarcZX10]

You were right: I accidentally pushed the unrelated diagnostics commit to this branch. I restored the intended rebased trace work; the current head is d1afc595e, and GitHub now shows the expected seven AGC draw-cache files. Sorry for the confusion.

I also corrected the PR description: this version uses the existing APS5_TRACE_DRAW_CACHE report and does not add a second trace switch or stderr path. The focused target builds and conventions pass. The CTest executable is blocked before its test body on this Xeon E5-1620 v2 due to missing LZCNT, BMI1, BMI2, MOVBE, AVX2, and FMA; I have updated the validation text to say that accurately.

[MarcZX10]

Rebased onto current main and pushed as 4f142b189. The update reuses traceDrawCache() / APS5_TRACE_DRAW_CACHE and the existing 10-second [draw-cache] report. It adds the stable-register, user-data, auxiliary-register, device, and shader-set attribution (plus an unchanged-component counter) beyond the existing new-register vs. user-word/pointer-only classification. The testShaderSetKey regression remains. Local conventions and git diff --check upstream/main...HEAD pass. I could not execute the focused test body on my Xeon E5-1620 v2 due to unsupported required CPU features; the previously reported supported-host run of the component test remains relevant.

[fma-rc]

I built 4f142b1 on main 8cd937e and ran agc_driver_register_key, agc_driver_draw_cache_reuse, agc_driver_draw_cache_reuse_off and agc_driver_draw_cache_reuse_verify: all 4 pass on an RTX 5080, on RADV (Radeon 610M) and on lavapipe.

I also ran Astro's Playroom (PPSA01325) into its heavy gameplay scene with APS5_TRACE_DRAW_CACHE=1 for 264 s with no throw. Three consecutive 10 s reports:

lookups	hits	no entry	stable registers	user data	auxiliary	unchanged
241848	11032	230226	4608	230226	1415	0
241706	9042	232404	4354	232404	1659	0
238983	8592	230249	4154	230249	1599	0

So every register-key miss there has changed user data, and only about 2% also changed stable registers. That is what the data-only path work in PR 5656 and PR 5609 goes after, and this report makes it easy to see.

[MarcZX10]

I believe the requested updates are now present in head 4f142b189: it builds on current main, reuses traceDrawCache() / APS5_TRACE_DRAW_CACHE and the existing report, and adds the per-component stable-register, user-data, auxiliary-register, device and shader-set attribution beyond the existing register/user-data split. The component-slot collision is fixed and the distinct shader-set regression remains.

For validation, fma-rc reports agc_driver_register_key, agc_driver_draw_cache_reuse, agc_driver_draw_cache_reuse_off and agc_driver_draw_cache_reuse_verify passing on RTX 5080, RADV Radeon 610M and lavapipe, plus a 264-second Astro's Playroom trace without throws. The hosted eligibility, conventions and report checks pass; hardware/path-filtered jobs skip. I still cannot enter the local test body on this Xeon because of the host CPU guard. Could you re-review the current head against your requested changes?

[adriadam10]

Re-checked the rebased head: it now reuses traceDrawCache() and the existing 10 s [draw-cache] report, components[4] carries the shader-set key and components[3] the device serial. agc_driver_register_key and the four agc_driver_draw_cache_reuse* tests pass with APS5_VERIFY_REGISTER_KEYS=1, and the new test does not build without the change. Default behaviour is unchanged with the variable unset. Confirmed by @fma-rc on RTX 5080, RADV and lavapipe.

Added on top as a small commit: dropped the unused identity temporary in drawRegisterKey (DrawRegisterKey.cpp) and a stray blank line. Merging once the full Build is green.