fix(agc): allow chunked imports as vertex and index buffers

[JJenkx]

What

Linux shared ranges fitting inside one 64 MiB udmabuf chunk can be bound directly as vertex or index input, but ChunkImportUsage omits both required usage flags. Override 2 (PPSA01907 v01.007.000) reports VUID-vkCmdBindVertexBuffers-pBuffers-00627 and VUID-vkCmdBindIndexBuffer-buffer-08784 with Vulkan validation.

Add VK_BUFFER_USAGE_VERTEX_BUFFER_BIT and VK_BUFFER_USAGE_INDEX_BUFFER_BIT to the existing usage mask, matching the ordinary host-pointer and sparse import paths. These are required by vertex-buffer binding and index-buffer binding. No allocation policy or draw ordering changes.

The regression imports an 8 MiB shared range, checks it takes the non-sparse dma-buf path, and checks the actual buffer creation flags admit both input roles.

Tested

Latest refresh: one linear topic commit 7ba8642a436deb6f0435a84e2beb4b705943ece4 on main 66e71b63d16eb07b940892ed2ad02afd625f7b3f, after #5390 merged. The conflict resolution retains upstream SparseBlockBytes and unaligned shared-range import support alongside this PR's vertex/index usage flags. All upstream device-replacement and unaligned-import tests remain, together with the existing 8 MiB small-import usage regression. Release libSceAgcDriver and the three test executables build. agc_host_import_sparse, agc_driver_graphics and agc_driver_sparse_dma_buf_import pass 3/3 on RADV and 3/3 on lavapipe, zero skips and zero core-validation errors/VUIDs (3.85 s / 3.31 s). Vulkan required, shader cache disabled. The host-import fixture uses mocked Vulkan with real Linux shared backing; lavapipe's real-device unaligned-range case reports a host-pointer import rather than sparse lead binding. Conventions, whitespace and a local merge-tree check pass. No fresh native title, full-suite, Windows or performance claim. Pinned report, RADV checks, lavapipe checks.

Historical validation below retains its original source revisions.

Head b3a58676bf108f830731f571975b810ef44b9dbe, one linear topic commit on main 93e54cedd3dd49eabf1c589f020b54758e0b6461. Linux 7.2.9-arch1-1, GCC 16.2.1 Release, Ryzen 9 7950X, RX 6650 XT / RADV Mesa 26.2.4-arch1.1; lavapipe Mesa 26.2.4 / LLVM 23.1.1.

• Full libs and the three selected test executables build. agc_host_import_sparse, agc_driver_graphics, and agc_driver_sparse_dma_buf_import: 3/3 pass on RADV and 3/3 on lavapipe, no skips (4.42 s / 4.12 s). The first is a mocked Vulkan device with real Linux udmabuf backing, not a hardware shader test.
• Negative control on main d9b696ddb with only the test added fails both missing-usage assertions. After adding the production fix, the exact same fixture executable passes; SHA-256 06618cfd9aed0f0e73f3efbcf1d8522bcdd8f7d07228309624969061985243b5.
• Native Override 2, installed RADV, pipelined draws enabled, Khronos validation 1.4.363 verified in both process mappings: baseline reports both VUIDs; adding only this fix reports neither. Both reach the menu and close normally (exit 0), restoring temporary settings. Native validation uses main d9b696ddb; the rebase onto 93e54cedd adds only upstream test guards and keeps the production file byte-identical. Baseline also has inactive capture diagnostics and tools-only replay changes, none modifying host imports.
• Separate CMASK host-stage and sampled-sRGB-view VUIDs remain, covered by fix(agc): include the host stage in the CMASK clear's host-access barriers #6781 and fix(agc): restrict sampled storage texture view usage #862 respectively; neither is a dependency of this fix. Dark 3D output remains unresolved. No visual fix, FPS gain, Windows validation or newly playable title is claimed.
• Conventions and whitespace pass. Full hosted builds still need collaborator dispatch.

Sanitized validation and screenshot: chunk-import-usage-validation-public.json, chunk-import-usage-menu.png, chunk-usage-main93-radv-tests.log, chunk-usage-main93-lavapipe-tests.log. No raw guest code, assets or full title logs published.

Separate combined diagnostic: testing/override2-api-fixes-main93 adds lvmino97’s #862 and XairRrR’s #6781 to this PR; its Override 2 startup/menu run closes normally with zero validation errors, while dark rendering persists. Further Recorder validation, recorder-complete-main93-validation-public.json, covers AnzouBeli’s #5883, Dextroid17’s #6155 and MarcZX100’s #6642 plus a separate CoverPass fixture correction: the full Recorder flow passes with zero validation errors in dynamic and legacy modes on both RADV and lavapipe. These are separate integration/test changes, not dependencies or a visual-fix claim for this standalone import PR.

Additional title coverage: Infliction: Extended Cut (PPSA02688) reports both missing vertex/index usage VUIDs on the local main 84b0d2f248 diagnostic stack. A combined integration on main f51724078f adds this PR plus #6782, #6781 and #862; both buffer VUIDs and the independent-blend, sampled-sRGB-view and host-stage VUIDs are absent through gameplay and normal shutdown. One pre-existing shutdown leak report remains. This is combined integration evidence, not an isolated A/B for this PR; dark lighting remains unresolved. Sanitized protocol and results.

Crimson Spires integration validation (2026-10-19 UTC)

A combined integration on main 3b20a366caa8a8de5509e7add433c55275aa6f29, head cdcb1b9b55630aaf7db1164656660afb9500a2d5, includes #7681, #7627, #7630, #6959/#6934, #6913, #862 and #6782. Full libs builds and 13 focused checks pass on RX 6650 XT / RADV Mesa 26.2.4-arch1.1, Arch Linux, with required Vulkan and core validation, zero skips (7.44 s total).

The warm-cache PPSA03568 native run completed after 307.859 s with exit 0. It recorded zero Vulkan validation errors; two warning types remain (image subresource layer-count compatibility and shader output not consumed, ten occurrences each). The 295 s capture shows the car gameplay scene with visible road and dashboard; the user independently confirmed both menu and gameplay rendering looked good. This is combined-integration evidence, not an isolated result for this PR, a PS5 image-match claim or a performance comparison.

Sanitized pinned report · Gameplay capture

Checklist

• Based on current main; no other open PR found adding these flags to ChunkImportUsage
• One topic per PR; follow-ups separate
• No new code comments
• No silent stubs or new unimplemented paths
• No investigation reports, game assets or images committed
• No third-party code added
• Generic Linux import behavior, not title-specific
• Depends on: none
• Unblocks: none (corrects Vulkan validation errors observed in PPSA01907; no bring-up stop or visible defect claimed fixed)
• GPUs tested: AMD Radeon RX 6650 XT / RADV Mesa 26.2.4-arch1.1 and lavapipe Mesa 26.2.4 / LLVM 23.1.1, Linux 7.2.9-arch1-1
• AI-assisted: yes

[github-actions]

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 1191	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 1324	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 4448	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 4449	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 5021	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6146	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6177	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6185	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6186	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6197	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6292	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6548	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6615	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6636	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6947	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6955	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 7201	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/HostImportSparseDmaBuf.cpp
PR 7233	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 7292	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 7332	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/HostImportSparseDmaBuf.cpp
PR 7446	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 7457	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 7473	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 7514	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 7642	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 7647	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp

[JJenkx]

Additional native evidence: Crimson Spires PPSA03568, opening-car dialogue, Linux / RX 6650 XT / RADV Mesa 26.2.4-arch1.1, default Khronos core validation enabled.

Baseline 355de8bdfcf25fb2af95ae9884c4163b10ac640d is main c4c2b3aa7371ee988c94ca62f1431ebaaf0688c8 plus the depth-bounds and scheduled-jump-table bring-up fixes (#7627 and #7630). It emits 10 validation headers each for VUID-vkCmdBindVertexBuffers-pBuffers-00627 and VUID-vkCmdBindIndexBuffer-buffer-08784: the chunk-import buffer lacks vertex/index usage. These counts are capped emitted messages, not a count of all invalid calls.

A local integration at bc8fdb728ff348c92fe5a65c0eefa80d03d097a7 adds this exact topic plus #6782 (independent blend), #6781 (CMASK host barrier stages), and #862 (sampled image-view usage). Both buffer VUIDs disappear; the other four baseline error categories also disappear. The two pre-existing warning categories remain. This is combined-integration evidence, not an isolated native A/B of #6913.

Both runs use a 1536x864 window, the same input schedule, APS5_HOST_IMPORT_MIB=16384, APS5_DRAW_CACHE_ENTRIES=4096, APS5_PIPELINED_DRAWS=1, and APS5_PLAYGO_UNLISTED_CHUNKS=installed, with retained caches. Both reach the same car dialogue and close normally at about 233 seconds; both PIDs are gone and the temporary udmabuf limit is restored to 64 MiB. The dark/malformed 3D scene persists. No visual fix or speedup is claimed.

Integration full libraries and affected fixtures build; agc_driver_graphics, agc_sampled_storage_base_level_usage, agc_host_import_sparse, and agc_driver_sparse_dma_buf_import pass 4/4 without skips (4.11 seconds). The separate CMASK draw/eliminate fixture also passes. Existing isolated buffer regression evidence above remains the evidence specific to this topic.

AI-assisted validation/reporting: yes.