fix(agc): copy a shader, context or user config register to memory with COPY_DATA

[LeandroLP23]

What

PS5 FSR4 Showcase (PPSA99011) stops with AGC driver: COPY_DATA at DWORD 69: COPY_DATA register, GDS or reference-clock source is not implemented once #5163 gets it past sceAgcInit. Right after a DISPATCH_DIRECT it copies three SH registers to memory with COPY_DATA (control 0x00100200: SRC_SEL 0, register; DST_SEL 2, memory):

• 0x2e0c: COMPUTE_PGM_LO
• 0x2e07: COMPUTE_NUM_THREAD_X
• 0x2e00: COMPUTE_DISPATCH_INITIATOR

Changes:

• Register source. COPY_DATA with a register source and a memory destination reads the value the queue holds. Shader registers are at 0x2c00, context registers at 0xa000 and user config registers at 0xc000. COUNT_SEL copies two consecutive registers.
  • A register outside those three banks is rejected at validation.
  • A register the queue has no value for throws at execution, so the copy never stores a guessed value.
  • ResolveStore leaves such a copy to Execute: the value depends on the registers at the packet's place in the stream.
• Dispatch registers. The CP writes the dispatch initiator into COMPUTE_DISPATCH_INITIATOR, and the group counts of a DISPATCH_DIRECT into COMPUTE_DIM_X/Y/Z. After each dispatch, Pm4::NoteDispatchRegisters does the same in the queue's shader registers. It runs in both the worker and the dispatch-ahead pass.
  • A DISPATCH_INDIRECT clears the three counts, because they come from memory, so a later copy of them throws.
  • Dispatch keys read only 0x207-0x213 from the shader registers, so they are unchanged.

Copies to a register, and GDS and reference-clock sources, are rejected as before.

Tested

Windows 11 (26300), AMD Radeon 860M (RDNA 3.5, driver 32.0.13058.8001), WinLibs GCC 15.2.0, Release, on main a0f5916a2.

• agc_driver_pm4 (extended) passes. It covers:
  • one and two shader registers, and a context register;
  • the initiator and group counts after a direct dispatch, and the initiator after an indirect one;
  • the throws for an unset register, and for the counts after an indirect dispatch;
  • the rejections outside the banks, including a two-register copy that leaves the shader bank;
  • that ResolveStore does not resolve a register copy.
• tools/check_conventions.py passes.
• PPSA99011 01.000.010 (catalog build, relinked with --windows --trap-sdk-syscall-stubs), with this branch and fix(agc): return 0 for out-of-range sceAgcInit without table lookup #5163 merged locally, 3 runs of 60 s:
  • without this change, every run stops at the COPY_DATA throw above, 7-10 s in;
  • with it, no run stops there. The title logs FSR4_SHOWCASE_SUBMIT probe_ms=0.61 and reaches FSR4_SHOWCASE_CHAPTER index=0 title=FSR 4 on PlayStation 5.
  • It then stops at later, separate limits: 2 of 3 runs at MUBUF resource descriptor register range overflow, and 1 of 3 runs with its own vkQueueSubmit(...) = -4.
• ctest -R '^agc_' -j 1 on this machine, test shader cache emptied: 480 passed, 2 skipped, and 1 failed, agc_driver_layered_volume_vertex_no_gpl. That test fails on main on this driver too: GPL-off pipelines on it, AMD Windows: identical AGC executables invert GPL pass/fail results after relocation #4768.
• Fork CI (lavapipe), run 38465741353: Linux 1078 of 1079 and Windows 1096 of 1097 pass, and the relinker jobs pass on all three hosts. The one failure in each job is conventions_ci_gate, which fails on any fork.

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 under Silent stubs in their library's technical debt file
• 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 (the title reaches this stop only with fix(agc): return 0 for out-of-range sceAgcInit without table lookup #5163)
• Unblocks: PPSA99011 sceAgcInit throws for an out-of-range version; PS5 FSR4 Showcase calls it with a pointer and expects an error result #4808, after fix(agc): return 0 for out-of-range sceAgcInit without table lookup #5163
• GPUs tested: AMD Radeon 860M (RDNA 3.5), driver 32.0.13058.8001, Windows 11
• AI-assisted: yes (Claude Code)

[xXPBSXx]

Windows test of #5477: Windows 10 (19045), MinGW-w64 GCC 15.2, Release, ninja -j 4, ctest called directly. Tree: main 7a0c838ca + merge of 5e36c08748c6cf464521faac41b24caddb4f6c32 (local merge 9fa5d9168), target agc_driver_pm4_tests.

• agc_driver_pm4: exit code 0, 3 of 3 runs.
• agc_driver_pm4_failure: exit code 0, 3 of 3 runs.
  Not tested: that the new Pm4 test cases fail without the change (no baseline run), Linux, a title run (PPSA99011 not available here), AMD, other ctest targets. These tests select no GPU.

AI-assisted (Claude Code).

[adriadam10]

Rebased onto current main (conflict in Pm4.cpp with the CopyDataSource/CopyDataBytes refactor; the register source now goes through those helpers) and added the missing technical-debt line in docs/dev/debt/libSceAgcDriver.md for COPY_DATA from a register and the dispatch registers it reads.

Checked: agc_driver_pm4 and the libSceAgcDriver scope pass on the rebase (only agc_driver_tessellation segfaults, as on main); the new test no longer builds without the change; COMPUTE_DISPATCH_INITIATOR/COMPUTE_DIM_* (0x2e00-0x2e03) are outside DrawKeyRegisters, so draw keys are unchanged. Merging once the full build is green.