[milescoviello]

What

Forspoken (PPSA03984) stops 43 s after start on main:

terminate called after throwing an instance of 'std::runtime_error'
  what():  DB_RENDER_CONTROL depth copy, resummarize or decompress is unsupported (register 0x0 = 0x000000c0)

0xc0 is DEPTH_COMPRESS_DISABLE (bit 6) and COPY_CENTROID (bit 7), with DEPTH_COPY and STENCIL_COPY (bits 2 and 3) clear (Mesa's GFX10.3 register definitions). COPY_CENTROID and COPY_SAMPLE (bits 8–11) select the sample a depth or stencil copy reads; AMD's PAL sets them only together with a copy (gfx9DepthStencilView.cpp), and #5499 uses them that way to pick the copied sample. On a draw with neither copy bit set, nothing is copied and they have no effect. DepthMaintenanceRejection and decodeDepth both rejected them as if they were a copy.

Both checks now accept bits 7–11. A draw with DEPTH_COPY or STENCIL_COPY is still rejected, whatever sample it selects.

depthMaintenanceTests gains 0x80, 0xc0 (Forspoken's value), 0x100, 0xf00 and 0x2fe3 as ordinary controls, and keeps a copy with a sample selection (0x84, 0x108, 0xf88) as maintenance. The clear test decodes a combined depth and stencil clear with 0xc0 and with 0xf00 added and requires the same clear, and still rejects a clear with a copy and a sample selection. On main's State.cpp the test stops at "ordinary depth controls were mistaken for maintenance"; with only DepthMaintenanceRejection changed it stops in decodeDepth on 0xe3.

Tested

Ubuntu 24.04, GCC 13.3 (CI's compiler), Release with CI's configure flags, on main e34c479, RTX 4070 Ti with NVIDIA 595.99.02:

• All 545 tests in libSceAgcDriver pass. agc_driver_stencil_ref_export is skipped because the device lacks VK_EXT_shader_stencil_export, and agc_host_import_sparse and agc_driver_sparse_dma_buf_import because /dev/udmabuf is not accessible here.
• Forspoken (PPSA03984), relinked with --to-intel and run on Linux. The two runs differ only in libSceAgcDriver.prx; main ran first with an empty shader cache, this PR second with the entries main's run had written:
  • main: stops at 43 s with the error above.
  • This PR: no DB_RENDER_CONTROL rejection; it was still running at 99 s, when it reached the 12 GB memory cap I set for the run. That is a separate problem: started from a filled shader cache, main reaches a 15 GB cap within 10 s as well, and a jemalloc heap profile of that start puts most of the memory in SPIRV-Tools optimization.
• @Dextroid17 checked it independently on Linux at head 573bf5a merged onto main a08de68 (default unoptimized build, mock Vulkan device): agc_driver_graphics passes 3 of 3, and with only Graphics/src/State.cpp put back to main it fails 3 of 3 at ordinary depth controls were mistaken for maintenance.
• @adriadam10 checked the DB_RENDER_CONTROL masks against the GFX10.3 register layout (0x101c in decodeDepth, ~0x2fe3 in DepthMaintenanceRejection): copy and resummarize/decompress still throw, and a sample selection without a copy is ordinary. agc_driver_graphics passed, and failed with main's State.cpp at ordinary depth controls were mistaken for maintenance.
• tools/check_conventions.py --base upstream/main passes, and tools/pr_overlap.py --branch finds no overlap.
• Forspoken (PPSA03984, PPSA03984-app0) on current main 8856f18, 1920x1080, pad input, one run each: main stops at 22 s on the DB_RENDER_CONTROL check (fix(agc): accept a copy's sample selection on a draw that copies nothing #5937); with fix(agc): accept a copy's sample selection on a draw that copies nothing #5937 it stops at 61 s on the indirect vertex buffer read (fix(agc): copy an indirect draw's vertex buffer only up to its last element #5975); with fix(agc): accept a copy's sample selection on a draw that copies nothing #5937 and fix(agc): copy an indirect draw's vertex buffer only up to its last element #5975 at 116 s on the depth pyramid view (fix(agc): treat a view past a depth surface's first level as memory #5982); with all three it draws an in-game scene by 60 s (the lower half still black, about 4 FPS) and runs to 195 s, where vkAllocateMemory for a render target fails with VK_ERROR_OUT_OF_DEVICE_MEMORY on the 12 GB RTX 4070 Ti. The memory goes to retired depth surfaces that are never freed. With fix(agc): free retired depth surfaces nobody samples #6465 added as well, it reaches a 600 s limit (on main 0c4816f).

Not checked on Windows or on a console.

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
• Unblocks: none
• AI-assisted: yes

🤖 Generated with Claude Code

[Dextroid17]

I ran this PR's test on Linux.

Setup: this PR's diff applied to main at 8cd937e (the PR's base is 421 commits behind it), default CMake build (CMAKE_BUILD_TYPE empty, unoptimized), clean environment. ctest -R '^agc_driver_graphics$', 3 runs each. The test did not open a GPU device, so this is a CPU-only run.

• with this PR: 3 of 3 pass.
• with Graphics/src/State.cpp put back to main's version (the new test kept): 3 of 3 fail at "AGC graphics: ordinary depth controls were mistaken for maintenance".

So the new test catches the old behaviour here. I did not check Windows, macOS or a real GPU.

Note: #5855 also edits Graphics/src/State.cpp and the same DepthStencilTests area of tests/Graphics.cpp, so whichever of the two merges second will likely need a rebase.

AI-assisted: yes

[milescoviello]

Thanks for running it. I've rebased the branch onto main 760ffa6; there were no conflicts, agc_driver_graphics passes on an RTX 4070 Ti, and tools/check_conventions.py passes. Noted about #5855 — whichever lands second gets a rebase.

AI-assisted: yes

[Dextroid17]

Thanks. I reran the same Linux check on the current head 573bf5a, after the latest push. Its change has the same git patch-id --stable as the head I tested before (802ff5d), so only the line numbers moved.

Setup: head 573bf5a merged into main at a08de68 (no conflicts; main has 14 commits that the head does not), default CMake build (CMAKE_BUILD_TYPE empty, unoptimized), clean environment (env -i, only HOME, PATH and VK_ICD_FILENAMES set). ctest -R '^agc_driver_graphics$', 3 runs each. CPU-only as before (the test uses a mock Vulkan device).

• with this PR: 3 of 3 pass.
• with Graphics/src/State.cpp put back to main's version (the PR's test changes kept): 3 of 3 fail at "AGC graphics: ordinary depth controls were mistaken for maintenance".

So the result from my first run holds on this head. I did not check Windows, macOS or a real GPU.

AI-assisted: yes

[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
PR 6146	core/libs/prx/libSceAgcDriver/Graphics/src/State.cpp L258-261, L918
PR 6762	core/libs/prx/libSceAgcDriver/Graphics/src/State.cpp L258-259

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 1395	core/libs/prx/libSceAgcDriver/Graphics/src/State.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 4040	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 4650	core/libs/prx/libSceAgcDriver/Graphics/src/State.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 5441	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 5499	core/libs/prx/libSceAgcDriver/Graphics/src/State.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 5500	core/libs/prx/libSceAgcDriver/Graphics/src/State.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 5925	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6188	core/libs/prx/libSceAgcDriver/Graphics/src/State.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6222	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6235	core/libs/prx/libSceAgcDriver/Graphics/src/State.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6312	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6320	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6481	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6523	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp

[adriadam10]

Checked the DB_RENDER_CONTROL masks against the GFX10.3 register layout (0x101c in decodeDepth, ~0x2fe3 in DepthMaintenanceRejection): copy and resummarize/decompress still throw, a sample selection without a copy is ordinary. agc_driver_graphics passes, and fails on main's State.cpp at "ordinary depth controls were mistaken for maintenance". Merging once the full Build is green.

[milescoviello]

Thanks for checking the masks. One thing that will hold up the Build: main at 65d56a2 doesn't compile on Linux with libstdc++. #5875's journal guards MeshConfiguration at 36 bytes, and #5884 made it 40, so PipelineJournal.cpp stops on its static_asserts. #6784 fixes that (MarcZX100 approved it). Until it lands, the Linux job here fails for that reason, not this change.

AI-assisted: yes

[milescoviello]

An update for the Build: #6784 is in, so main compiles on Linux again. One test still fails on main itself: agc_driver_graphics, since #5804 (#6792, a zero-filled vertex code span). #6789 fixes it and is approved, so a Build of this PR before #6789 lands will show that one failure regardless of this change. With #6789 applied, agc_driver_graphics passes with this PR, as you saw.

AI-assisted: yes

[adriadam10]

Rechecked on current main 93e54ce: head merges cleanly, agc_driver_graphics passes and still fails on main's State.cpp; merging once the full Build is green.