perf(agc): handle the dword pattern-fill kernel as a buffer fill

[orbz22]

What

Astro's Playroom (PPSA01325) runs a 64-thread compute kernel (69 dwords) every frame to fill DCC keys, HTILE and other metadata. It ran as a shader. Before the fill, every key read of those surfaces waited for a drain (on the UI target 0x534830000, among others).

What the kernel does, read from its code and confirmed by the new test against the translated kernel:

• It covers every i below min(launched threads, s8, num_records). The group id is in s10, so the kernel has exactly 10 user SGPRs.
• Into dword i of a 4-byte-stride buffer (s[0:3]) it stores pattern[min(i % s9, 3)], where the pattern is s4..s7 and s9 is the period.
• The modulo is computed with a float reciprocal. Remainders 0, 1 and 2 select s4, s5 and s6; anything else stores s7.

MatchDwordPatternFill (new header Driver/Memory/DwordPatternFill.hpp) recognizes the exact code and accepts only one shape:

• a direct, non-partial 1-D dispatch;
• a 4-byte-stride, unswizzled, non-add-tid buffer;
• a period of 1, 2 or 4 dwords;
• whole 16-byte units into a 16-byte aligned destination.

That shape is a 16-byte pattern fill, and fillBuffer takes it the same way it takes the existing fill kernel: clear path, FillBuffer, key memo. Everything else still runs the kernel. APS5_NO_FILL_HLE turns both matchers off.

#1335 recognizes a different, typed pattern-fill kernel (18 dwords, pattern and control read from buffers). Both PRs touch fillBuffer, so whichever merges second needs a small rebase.

Tested

Windows 11, WinLibs GCC 15.2, RTX 3060 Laptop:

• New agc_driver_dword_pattern_fill_tests (tests/execution/DwordPatternFill.cpp):
  • the matcher accepts the kernel, and refuses every single-bit change of its code, a truncated copy, periods other than 1/2/4, partial units, unaligned destinations, add-tid descriptors, other user SGPR counts, partial dispatches and a missing group id;
  • it clamps to the descriptor's records;
  • ten cases run through Submit: matched ones and unmatched ones (periods 3 and 5, a count of 250, a partial dispatch).
• The test passes as is. With APS5_NO_FILL_HLE=1, the translated kernel meets the same expectations, which is where the min(i % s9, 3) reading comes from.
• agc_driver_tests and agc_driver_pm4_tests pass, and tools/check_conventions.py --base upstream/main passes.
• PPSA01325, A/B on the title screen with this commit as the only difference: 12.68 -> 13.96 FPS (78.9 -> 71.6 ms per frame), mean of three runs per side, B ahead in every pair. Current main does not reach the title, so both sides ran on an integration branch; details and per-run numbers are in the comments.

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
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed in TechnicalDebt
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• AI-assisted: yes (Claude Code)

🤖 Generated with Claude Code

[oneandonlydean]

Could you measure main against main plus this PR alone on PPSA01325, frame time on the same scene over a few runs? Perf PRs need numbers for the change itself (#218); the 40-57 ms is the cost on main, and the after run had other changes in it.

[orbz22]

@oneandonlydean measured as an A/B with only this change between the two sides. PPSA01325 title screen, the same scene in every run; frames counted over 111-202 s of each run, three runs per side, A and B alternating.

Current main does not reach the title in PPSA01325. It stops in turn on:

• missing static shader ABI register 201 (fix(agc): defer shaders that cannot be prepared at registration to their first use #1836);
• rectangle ABI refers to a replaced shader header (fix(agc): accept bound copies of a registered shader header that differ only in user_data #1837);
• DB_RENDER_CONTROL depth clear draws (feat(agc): add experimental depth clear and copy recognition #1395);
• vertex fetch exceeds descriptor record count (no PR yet);
• the null pixel program (fix(agc): prepare the null pixel program at draw when its variant was not prepared #1902);
• rect-list rectangle artifacts (fix(agc): prepare a rect-list draw's missing rectangle at draw from its prepared stages #1928);
• a read through a null SRT pointer at +0x28 (feat(agc): implement image and sampler tables indexed at runtime offsets #757-feat(agc): implement deferred faults for reads through null SRT pointers #759, which no longer apply to current main).

So both sides run on my integration branch: main from 2026-10-08 plus those PRs, with the same build and the same switches on both sides. The only difference is the fill kernel:

• A: the branch's earlier local matcher for this kernel is turned off (APS5_NO_DWORD_FILL_HLE=1), so the kernel runs as a dispatch, as on main;
• B: this PR's commit cherry-picked on top (it replaces that local matcher).

run	A (kernel dispatched)	B (this PR)
1	13.93 FPS, 71.8 ms	14.80 FPS, 67.6 ms
2	11.90 FPS, 84.0 ms	13.96 FPS, 71.7 ms
3	12.22 FPS, 81.8 ms	13.12 FPS, 76.2 ms
mean	12.68 FPS, 78.9 ms	13.96 FPS, 71.6 ms

About 7 ms per frame, with B ahead in every pair. That is far below the 40-57 ms in the description. Those were the cost of the waits on 0x534830000 measured on main, not an A/B, so I've replaced them in the description with these numbers.

[oneandonlydean]

Conflicts with main in BufferFill.cpp now that #1335 is in; please rebase. It matches a different kernel from #1335, so both stay.

[adriadam10]

Rebased onto current main as @oneandonlydean asked (conflict in fillBuffer with #1335; both kernels stay). The dword pattern-fill match is evaluated first and feeds the same clear / FillBuffer / key-memo path; MatchesPatternFillKernel and the 16-byte kernel are untouched. Checked: local build and ctest of libSceAgcDriver green (306/306), agc_driver_dword_pattern_fill and agc_driver_pattern_buffer_fill pass. Merging once CI is green on the rebased head.