[EagleBearerYS]

What

This is the follow-up test that thuanquang suggested on #7346 (comment), which has since merged. It adds the 32_SINT, 16_UINT and registered-allocation cases that #7346 left for a separate PR.

The case it covers: NHL 26 (PPSA26785) stopped at start-up on main with frankea's line from #7304 (main 922dc89, RTX 4090):

compute shader 0x1150101000: AGC graphics: 32-bit integer read of the depth-layout texture 0x11c10f0000, which is no depth surface drawn with, is not implemented

#7346 reads such a T# over guest memory through a float view, and frankea's run with it reaches the legal screen. #7346's own case reads a 128x128 32_UINT texture.

The test: agc_driver_depth_bits_guest_memory, in tests/execution/DepthBitsGuestMemory.cpp, the only file this PR adds:

• The surface: 256x192 in SW_64KB_Z_X, laid out with the driver's swizzle equations, with no depth surface drawn.
• The reads: 32_UINT, 32_SINT and 16_UINT T#s, each from untracked memory and from a registered guest allocation. On lavapipe the guest allocation's later reads go through a storage image (the GPU detile), so both the snapshot path and the storage-image path run.
• The check: every texel is compared bit for bit. The values include denormals (0x00000001, 0x007fffff, 0x80000001), quiet and signalling NaNs (0x7fc00000, 0x7f800001, 0x7fbfffff, 0xffffffff) and infinities.

Tested

• GitHub's Linux (lavapipe, Mesa 25.2.8) and Windows runners (verify run):
  • Red: with fix(agc): read a 32-bit depth-bits T# over guest memory through a float view #7346's ShaderResources.cpp change reverted, agc_driver_depth_bits_guest_memory and fix(agc): read a 32-bit depth-bits T# over guest memory through a float view #7346's agc_driver_depth_plane_sampling both fail on the Linux runner.
  • Green: with the change, both pass 10 of 10 runs there.
  • Windows: they build and skip there, as GitHub's Windows runner has no Vulkan device.
  • Suites: libkernel passes on both runners, and libc on Windows. libc's precise_wait timing test fails on the Linux runner with and without fix(agc): read a 32-bit depth-bits T# over guest memory through a float view #7346's change.
• Windows, the full CI build (fork Build): 1441 of 1441 tests pass, the new one skipping, and check_skips passes.
• Linux, a run that mirrors CI's Linux job (a cloud machine, Ubuntu 24.04, GCC 13.3, Mesa lavapipe 25.2.8 with LLVM 20.1.2, run as an unprivileged user):
  • Totals: 1414 tests on this branch, against 1413 on fix(agc): read a 32-bit depth-bits T# over guest memory through a float view #7346's head. The new test is the only difference, and both trees show the same 5 failures, the IPv6 tests that machine always fails. check_skips flags only its known IPv6 skip, guest_sce_net.
  • Red there: with fix(agc): read a 32-bit depth-bits T# over guest memory through a float view #7346's change reverted, both tests fail with "32-bit integer read of the depth-layout texture ..., which is no depth surface drawn with, is not implemented".
  • Green: both pass with the change, and the new test passes 20 of 20 repeats.
• Bits: on Mesa lavapipe 25.2.8 (llvmpipe, LLVM 20.1.2), every value keeps its bits, NaN and denormal patterns included.
• Not tested: AMD, NVIDIA and Intel hardware, and NHL 26 itself. The test changes no emulator code.

Checklist

• Based on current main (on fix(agc): read a 32-bit depth-bits T# over guest memory through a float view #7346's merged head 4aaa125, and it merges cleanly with main a86577e); 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 (none added)
• Generic behaviour, not specific to one title
• Depends on: none (fix(agc): read a 32-bit depth-bits T# over guest memory through a float view #7346 has merged)
• Fixes: none
• Unblocks: none; fix(agc): read a 32-bit depth-bits T# over guest memory through a float view #7346 unblocked PPSA26785 NHL 26 (PPSA26785) #7304
• GPUs tested: Mesa lavapipe 25.2.8 (llvmpipe, LLVM 20.1.2), Linux (GitHub's runner and a cloud machine)
• AI-assisted: yes