perf(agc): stage read and written buffer elements a dispatch sweeps

[oneandonlydean]

What

A written V# element bound in place in a host import makes every load a PCIe round trip. In one title's particle update, a kernel loads and stores a 10 MiB pool (2048x64 threads over 131072 records of 80 bytes). That is above the 2 MiB written window, so it stayed in place and took about 4.8 ms of GPU per frame.

Two commits:

1. feat(shader): DescriptorBinding gains bufferRead next to bufferWritten and bufferAtomic, filled from the resource tracker's existing per-buffer read flag. It is kept in the shader disk cache, which bumps the format to 24. No key or ABI change.
2. perf(agc): an element is now staged in device-local memory up to 16 MiB when all of these hold:
   • the shader also reads it;
   • it is written by a direct dispatch;
   • the dispatch's threads times the V# stride cover at least half of it.

The coverage test keeps a kernel that updates about 12k records of a 16 MiB pool in place: staging all of it would cost about 1.3 ms of copies for a 0.08 ms kernel. Write-only elements, indirect dispatches and raw (stride 0) buffers keep the 2 MiB window.

Tested

• Linux (CachyOS, x86-64, GCC, RTX 4090), Release with BUILD_TESTING=ON: ctest -R "agc|shader" passes 320/321. The one failure, agc_driver_dcc_keys_first_write, fails the same way on current main.
• agc_shader_disk_cache_tests: bufferRead survives the disk cache round trip and the binding plan.
• agc_driver_recorder_tests, new readWrittenStagingTests, run on its own on the RTX 4090 (the full binary stops earlier on main with a 60 s lease wait, unrelated):
  • a read and written 4 MiB element is staged;
  • the same element written only stays in place;
  • a dispatch touching 16 KiB of it, or of unknown size, keeps it in place;
  • copy-backs leave the guest bytes unchanged.
• Astro Bot (PPSA21564), measured on our fork, which carries the same change: the particle update 0x500779a00 went from about 4.8 ms of GPU per frame to 0.06 ms plus about 0.9 ms of copies.

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)

[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
#757	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp L44 core/shader/recompiler/ShaderDiskCache.cpp L68-70 core/shader/recompiler/ShaderDiskCache.hpp L15
#758	core/libs/prx/libSceAgcDriver/tests/ShaderDiskCache.cpp L44 core/shader/recompiler/ShaderDiskCache.cpp L68-70 core/shader/recompiler/ShaderDiskCache.hpp L15
#822	core/shader/recompiler/ShaderDiskCache.hpp L15