fix(agc): preserve CPU writes after resident linear target draws - #8246

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.

I marked this PR draft after a second correctness review. Deferring writeback alone is insufficient: the current write-watch policy preserves whole 64 KiB blocks, which can discard the draw output in the same block as a post-draw CPU store. Also, a writeback recorded during a completion needs to execute synchronously rather than remain in a newly opened recorder batch. I am working through a region-aware writeback path and will update the PR when the regression and completion ordering are addressed.

Update after checking the reporter’s run on current main: my earlier proposed keep-block explanation was wrong. The seed causes a re-upload after the CPU write, and the draw completion flush happens before later CPU writes. The unseeded, different-output case demonstrates the actual whole-target writeback ordering bug. I also found a lifecycle blocker in my attempted deferral: cache eviction marks a texture released, which makes FlushPending ignore it. I discarded that incomplete local attempt; the published branch remains draft while I work out a fix that keeps the pending image discoverable and orders its later writeback at the access that needs it.

I pushed two commits to #8246: fe29558e94 defers linear target eviction writeback and records the draw as a pending writer; 285884da27 updates the regression to flush and wait for readback before checking memory. The writeback keeps changed CPU bytes within the same 64 KiB block while storing the remaining GPU bytes. The focused target builds and links locally, conventions pass, and CTest is blocked before Vulkan initialization by this host CPU’s missing required extensions. I have kept the PR as a draft pending a run on a supported CPU/GPU.

Follow-up: I found the first byte-merge revision could compare against an upload baseline that was stale across recorded draws. Commit ec9d6fecb5 snapshots guest bytes immediately before each recorded linear draw and uses CPU-specific write stamps when preserving changed bytes. The focused target rebuilt and linked successfully; runtime remains blocked before Vulkan initialization by the host CPU feature guard. I have kept the PR draft.

Thanks for the detailed review and for checking the previous code on your RTX 5050. Since that review, the fix changed: it now leaves target results pending after batch completion and stores them on the later read/flush path; commit ec9d6fecb5 also refreshes the byte baseline before each draw, and 50f477fccc reuses the full guest snapshot rather than limiting byte merging to 1 MiB. I rebuilt the focused target, but this Xeon host is stopped by the CPU feature guard before Vulkan. If you have time, could you run the current draft head (50f477fccc) through agc_driver_linear_target_resident on your GPU? The PR remains draft while that runtime behavior is unverified here.

Sharing a tested alternative implementation and regression coverage for #8152 so we can coordinate with this PR.

Branch: rajvveer/AnyPS5 — fix/agc-linear-target-cpu-writes

• Implementation: 9787db292
• Queued-draw and cached-recipe coverage: e5b9ac546

This approach captures the actual image on the GPU before a recorded linear draw and merges only changed bytes into guest memory. Atomic word updates preserve untouched channels sharing a word. Image and guest baselines remain separate for subsequent refresh; CPU write-back and padded array layers are covered too. It keeps the recorded path.

The tests extend Celegans12's reproducer: seeded/unseeded partial draws, masked float/RGBA8 channels, partial geometry, later blending, queued overlapping draws, actual cached-recipe replay, and GPU retile cases for 1/2/4/8/16-byte elements. Recipe replay is unseeded so a deferred re-upload cannot mask the original bug.

Validation: Windows 11 Pro 26200, WinLibs GCC 15.2.0 r7, Release with SPIRV-Tools; NVIDIA RTX 5070 (617.42) and AMD Radeon(TM) Graphics (32.0.21030.2001).

• Independently built base 3fd1cc393: the compatible regression overlay reproduces the overwrite on both GPUs; synchronous controls pass.
• 36/36 fresh focused runs passed at e5b9ac546, with the correct devices and Khronos 1.4.363 layer verified, synchronization validation enabled, and no validation errors/warnings or skips. This includes five modes plus actual recipe replay.
• 118/118 common selected tests per GPU passed on base and change. Existing broader-suite diagnostics matched by test name, message and count: NVIDIA 5 errors/10 warnings; AMD 6 errors/10 warnings. No new VUIDs; this is not a claim that the entire suite is validation-clean.
• Full runtime/relinker builds pass. The paired suite preceded the final recipe-only test tightening; production code was unchanged, and the final build/focused runs were repeated.

These results apply to the linked branch; I have not executed this PR's current 50f477fccc head. No real-title, Linux/macOS, console or performance result is claimed. The snapshot adds GPU work and temporary memory.

Would incorporating the implementation or regression coverage here be useful, or would you prefer a coordinated alternative PR? I have kept the separate PR on hold.

I chose to integrate your implementation and regression commits into #8246 rather than open a second PR for the same issue. Both commits remain authored by you (unknown <rajveershekhawat626@gmail.com> as recorded in Git), and the head is now e32a19b4d7957795a3b5cf21848b41889bd193a5, rebased on current main 35fa2bdd. The older-base GPU/build results and the rebase validation limits are now explicit in the PR description. Could you rerun the focused matrix on this exact head when possible? I will keep the draft open pending that current-head check.

Correction to my previous note: I said I would keep the PR draft open, but GitHub shows it was already marked ready for review at 07:46 UTC. I had the state wrong; I did not change the draft/ready setting as part of integrating your commits. The PR remains ready for review, with the current-head runtime rerun still outstanding.

The requested exact-head rerun is complete: e32a19b4d7957795a3b5cf21848b41889bd193a5 was rebuilt in a separate clean worktree and passed 36/36 focused executions. agc_driver_graphics passed too. Independently rebuilt main 35fa2bdd also passed the default resident control on both GPUs.

Hardware: Windows 11 Pro 26200, NVIDIA GeForce RTX 5070 / driver 617.42, and AMD Radeon(TM) Graphics / driver 32.0.21030.2001; GCC 15.2.0 Release with SPIRV-Tools.

The matrix was five resident modes (default, unwatched, draw transitions, eager write-back, synchronous linear targets) plus cached-recipe replay, three repetitions on each GPU. Every process confirmed the intended GPU and loaded Khronos validation layer, with ANYPS5_REQUIRE_VULKAN=1, ANYPS5_NO_SHADER_CACHE=1 and synchronization validation enabled. Zero VUIDs, synchronization hazards, validation errors/warnings or skips.

A small test-only follow-up is also available: 66f62d174 — strengthen linear target replay and array coverage. It changes only LinearTargetResident.cpp and applies cleanly to e32a19b4:

• Capture draws a quarter-width area; cached replay draws the original half-width area. The assertions now require pixels produced by replay, alongside preservation of the CPU-written right half.
• The padded-array fixture now also covers storage snapshot, merge and refresh in imported memory at layer 1, preserving the existing CPU fallback. Distinct initial layer contents and completion of the initial upload make an incorrect baseline-layer offset observable. Full-buffer and refresh checks cover GPU results, CPU writes, padding and the untouched layer.

The follow-up was separately built and tested at 66f62d174: 36/36 focused executions passed, followed by five successful checks after the negative-control experiment was restored and rebuilt (agc_driver_graphics, plus resident and recipe on both GPUs). Temporarily removing the per-layer baseline offset made the new imported-snapshot assertion fail on both GPUs, with the CPU fallback still passing and no validation diagnostics. That mutation is absent from the commit. Convention and whitespace checks pass.

Please cherry-pick 66f62d17479bf7b393326c48926dea6bea31c6f0 if you would like to include this test tightening. The two 36-run results above belong to their respective commits; the follow-up is not yet part of this PR. No real-title, performance, Linux/macOS runtime or fresh full-suite result is claimed.

I rebuilt exact head e32a19b4d7957795a3b5cf21848b41889bd193a5 on an RTX 4070 Laptop GPU. The existing resident and cached-recipe tests pass, but an additional concurrent-byte regression exposes lost CPU stores.

Attached is a tested correction extending Rajveer Shekhawat’s implementation and MarcZX10’s integration. The seeded/unseeded regression derives from Celegans12’s reproduction. I’m offering this for integration into the existing PR.

Confirmed failures and correction

The concurrent regression uses a 64 × 4096 RGBA8 target. The GPU writes only blue while a CPU thread updates counters in red/green and checks the preceding values. On e32a19b4, it detected 5,668 lost CPU updates, although blue was correct and later CPU sweeps repaired the final counters. Checking only the final buffer misses this failure.

GPU word-level compare-and-swap does not provide atomicity with ordinary CPU stores to adjacent bytes. The correction uses a separate publication shader with write-only uint8_t storage-buffer stores, without destination loads or atomics. The driver already enables storageBuffer8BitAccess; ordinary detile/image variants remain unchanged.

A second regression exposed a stale-baseline problem: an actual GPU write equal to stale image contents can be omitted by the before/after comparison. PrepareLinearDraw now settles prior results and overlapping aliases, then uploads guest contents before taking the GPU baseline. This regression fails with the byte-store-only intermediate and passes with the complete correction.

The comparison is not treated as a write-coverage detector. It depends on establishing a coherent baseline first. Unsynchronized CPU/GPU stores to the same byte still require ordering. Publication remains in the recorded command stream, with the existing reader/host barriers and resource retention. The new checks use the established WaitIdle boundary without an additional test-only flush.

Tested

Hardware: NVIDIA GeForce RTX 4070 Laptop GPU, NVIDIA 592.82 / Windows driver 32.0.15.9282, Windows 11 Pro 10.0.26300. WinLibs GCC 15.2.0 r7, Release, Ninja; Khronos validation layer 1.4.363.0.

• Rebuilt main 35fa2bdd824f1a77d5d70e7feeae85fd6d9baf04: 8 failures in 12 shared focused entries. The original unseeded regression preserves all 512 GPU pixels but overwrites 7,680/8,192 CPU-written bytes. Seeded and synchronous controls pass.
• Corrected local revision fea0c0cd8a2e8f85454e0d4b5b917915e0e3035b: 15/15 focused tests pass, including concurrent CPU updates, queued draws, channel masks, partial geometry, dependent GPU copy and CPU-visible label, cache eviction, aliases, same-value writes, expanded resident coverage and cached-recipe replay.
• After the main comparison, restored and rebuilt the correction: 15/15 pass again.
• Four key cases repeated five times each: 20/20 executions pass, including zero lost CPU updates in every concurrent run. Same-output seeded/unseeded, sequential CPU-writer, direct-transfer reader and synchronous controls also pass.
• Main uses the common runtime fixtures with identical assertions. The new same-value and expanded resident/recipe cases require APIs absent from main and are reported separately. Main’s original resident control passes; descriptor-capture mocks use the matching API version.

Every GPU run used:

ANYPS5_REQUIRE_VULKAN=1
ANYPS5_NO_SHADER_CACHE=1
ANYPS5_GPU=RTX 4070 Laptop
VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation
VK_LOADER_DEBUG=layer
VK_LAYER_ENABLES=VK_VALIDATION_FEATURE_ENABLE_SYNCHRONIZATION_VALIDATION_EXT

Actual validation-layer loading was verified. Base and correction logs contain zero VUIDs, synchronization hazards, missing-layer warnings or skips. The shared deprecated-setting notice is not counted as a new validation error.

Focused candidate command:

ctest --test-dir build-review-8152 -R '^agc_driver_(linear_target_(concurrent_bytes|cpu_writes.*|publication.*|resident|recipe)|graphics|alias_copy_label|array_layer_target)$' -V --timeout 30

Convention and whitespace checks pass. spirv-val --target-env vulkan1.1 accepts the publication shader; its compiled module has 8-bit storage, byte array stride 1 and no atomic operations.

The attachment is a plain unified diff against e32a19b4d7957795a3b5cf21848b41889bd193a5, without commit metadata. Apply it with git apply <attachment>.

SHA256: d7dbee013607048a06f05efea9b4ed3a4ab31086c1b396670e5ca4cf7b403bee

The additional upload/snapshot costs work and temporary memory. No game compatibility, performance improvement, Linux/macOS runtime, fresh full-suite or other-GPU result is claimed for this correction.

AI-assisted: yes.
8152-pr8246-followup.diff.txt

Integrated the offered correction into #8246 as commit 030409ad4a86041c29fa65bf018632452d5b5cab after rebasing all three PR commits onto current main 891d5675150936bbb330a1f27bacc86113dd2532. The new commit preserves embersas as its author; the two Rajveer commits also retain their original author metadata.

The contributor-reported RTX 4070 Laptop run and correction-tree test results are recorded in the PR description, explicitly as pre-latest-rebase results. On the rebased head, git diff --check passes. The local conventions script stops at its duplicate-export scan with E2BIG, and hosted metadata checks were pending at the last poll. I could not run a current-head Vulkan build/runtime on this host.

This is a runtime run of the exact current head e32a19b4d, which was still outstanding. I merged it onto main 02fb1ce53 (a clean merge, after #8172).

Setup: Windows 11 (26300), WinLibs GCC 15.2.0, Release, BUILD_TESTING=ON. The GPUs are an RTX 5070 Ti (NVIDIA 617.42) and an AMD Radeon iGPU (AMD proprietary 2.0.353). Both runs used ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1; the validation layer was not used.

• agc_driver_linear_target_resident passes 3 of 3 on each GPU. So do agc_storage_writeback_blocks, agc_write_tracking and agc_write_tracking_metadata_commit. The test logs linear target resident device: NVIDIA GeForce RTX 5070 Ti and ... AMD Radeon(TM) Graphics respectively.
• Negative control: the PR's test uses PrepareLinearDraw, FinishLinearDraw and the new Dispatch overload, so it doesn't compile on main alone. Instead I kept the whole PR and restored only main's Draw.cpp, which removes the PrepareLinearDraw and FinishLinearDraw hook-up. The test then fails 2 of 2 on each GPU with unseeded scissor: write-back overwrote the CPU's untouched right half in row 0.
• Related set of 75 tests (agc_driver_graphics, which includes the new TextureDetiler.cpp cases, plus storage, partially_resident, msaa, multisample, label, stencil, tessellation, write_tracking, resident and so on): 75 of 75 pass on each GPU. agc_driver_stencil_ref_export skips on NVIDIA because the GPU has no stencil export.

Not measured: the cost per draw. PrepareLinearDraw runs WriteBack() and then copies the whole image into a new 2x-size DeviceBuffer before every recorded draw into a resident linear target, and I have no title that exercises this path to time it.

AI-assisted: yes (Claude Code); results above were run on this machine.

I tested #8246 with Ghost of Yōtei (PPSA26344) on Windows 11 and an RTX 3080 Laptop, since #6892's resident linear targets are what broke our warm world runs in the first place. The unit-level fix holds up well on our tree. In the game, though, it doesn't fix the spawn: with resident linear targets on, Atsu still isn't at the river.

One caveat first: this isn't plain main. Our tree is main b509767 merged with our local Yōtei work (local/merge19 a5d6e17, where resident linear targets are opt-in behind APS5_RESIDENT_LINEAR=1). I applied your head 030409a on top as one diff. There were three small conflicts (Draw.cpp, Texture.cpp and the CMake test list), and I kept both sides in each. The shader cache version didn't change.

The tests all pass. Your new tests, the upstream resident-linear tests and the nearby storage and write-tracking tests went 22/22, three times each, with resident linear on. My #8152 reproducer passes 5/5 in both the seeded and the unseeded case. With only your Draw.cpp hooks taken out again, it fails 3/3 with 7,680 of the 8,192 right-half bytes rolled back, as before.

In the game, the warm run from our Ainu-village save never shows Atsu. From frame 078 on, the camera sits under the water looking up at the forest. That's the same family of symptoms as our earlier resident-linear runs (T-pose in the sky). With resident linear off, the same save spawns him at the river every time. The path does run there: in the world, about 15-43 recorded draws per 10 s go into resident linear targets, and each one gets your "linear draw baseline" upload.

I also wanted to know whether the game just reads those targets before the recorded write-back lands. So I added a local debug switch that does a full Recorder::Sync right after every recorded linear-target draw. That makes the write-back land before the next packet. It didn't help: the camera was still under the water. My guess is that the bytes reaching guest memory, or the resident image that the next draw starts from, are wrong in a case the tests don't cover, not that the order is wrong. I can't rule out an interaction with our local changes yet, though. Neither run lost the device or hit a GPU write fault.

If it helps, I can run an instrumented build next. For example, it could dump a linear target's guest bytes after one of these draws with resident linear on and off and diff them. Or I can try a newer head. Is there anything you'd want me to log in particular?

@Celegans12, thanks for offering. A targeted capture on the current head 030409ad4a would help. If convenient, pick one linear target associated with the first incorrect water frame and record its guest address, dimensions/pitch/format, the guest bytes immediately before the draw baseline upload, the resident image bytes after the draw, and the guest bytes after write-back. The key is to identify the first point where those byte ranges diverge; no need for broad per-draw logging. If selecting a target is unclear, a newer-head run alone would still tell us whether the behavior changed.

Thanks for the detailed title run and the sync control. That changes the next question: the isolated CPU-write regressions pass, but this title still diverges even when each writeback is forced to complete before the next packet, so completion ordering alone does not explain the spawn failure. The most useful next capture would be the first target/draw that differs between resident-linear on and off: guest address, dimensions/format/pitch, bytes before the draw, resident image bytes after it, bytes written back, and the next consumer’s read. A narrow capture around the first divergence should distinguish stale baseline/image contents from a format/layout or alias mismatch. If you can share such a trace (or an instrumented reproducer), I can investigate the next fix in this PR.

Thanks, this separates the #8152 byte-preservation case from the remaining Yōtei spawn failure: the focused CPU-write/repro tests and their negative control still validate the intended fix, while the title remains below water even with the per-draw sync switch. For the first affected target, could you log a compact identity (guest address, format, extent/pitch) and a hash plus a few sampled bytes at three points: after PrepareLinearDraw establishes the baseline, after the recorded draw completes just before its completion callback, and after WriteBack? That should show where the frame used for the spawn first diverges without dumping broad title memory. Comparing resident-linear off/on on the same saved state would help anchor it. There is no need to spend time timing the broken frame; I agree those scenes are not comparable.

Thanks for the detailed run. The result narrows this down: the reproducer confirms the CPU write-back regression is fixed, while forcing a full sync after each recorded draw does not fix the Yōtei scene, so I agree we should look at the bytes and resident image contents rather than add more ordering waits.

If you can run one instrumented comparison, please pick a target touched during frames 078–090 and log/dump:

• the target address/range, format, dimensions, pitch/tiling, and draw sequence number;
• its guest-memory bytes after that draw's write-back has completed;
• the same bytes immediately before the next draw that reads or writes the target; and
• the corresponding resident-image contents at those two points, if the existing debug path can read them back.

Please capture resident-on and resident-off runs from the same save and report the first differing byte/pixel with the owning draw. A small bounded set around one target is enough; no need for a full frame dump. That should tell us whether the mismatch is in the write-back, the resident image used by the next draw, or another local Yōtei change.

✅ NVIDIA GPU test: passed

PR head 8868994 merged onto main 9ab6a28 · tested 2026-11-04 22:15 UTC · 35 min

agc_driver_submit_progress: failed in 3 of the last 10 main runs
agc_submit_progress: failed in 4 of the last 10 main runs

agc_driver_image_table
guest_mmap_executable
guest_socket_options_multicast

agc_driver_linear_target_concurrent_bytes
agc_driver_linear_target_preserved_bytes
agc_driver_linear_target_publication

VUID-VkDeviceCreateInfo-pNext-pNext (554 tests): agc_driver, agc_driver_address_draw_space_reuse, agc_driver_address_passes, agc_driver_alias_copy_label, agc_driver_alpha_to_coverage, agc_driver_array_layer_target, agc_driver_bc_block_copy, agc_driver_bda_anchor_reads, and 546 more
VUID-VkShaderModuleCreateInfo-pCode-08737 (8 tests): agc_driver_dpp16_operands, agc_driver_dpp8_pk_fmac, agc_driver_dx10_clamp, agc_driver_float16_fma_rounding, agc_driver_float16_unused_nan_payloads, agc_driver_packed_alu, agc_driver_packed_f16_denormal_modes, agc_driver_packed_float16_native
VUID-VkShaderModuleCreateInfo-pCode-08739 (8 tests): agc_driver_dpp16_operands, agc_driver_dpp8_pk_fmac, agc_driver_dx10_clamp, agc_driver_float16_fma_rounding, agc_driver_float16_unused_nan_payloads, agc_driver_packed_alu, agc_driver_packed_f16_denormal_modes, agc_driver_packed_float16_native
VUID-VkShaderModuleCreateInfo-pCode-08741 (8 tests): agc_driver_dpp16_operands, agc_driver_dpp8_pk_fmac, agc_driver_dx10_clamp, agc_driver_float16_fma_rounding, agc_driver_float16_unused_nan_payloads, agc_driver_packed_alu, agc_driver_packed_f16_denormal_modes, agc_driver_packed_float16_native
VUID-vkCmdCopyImage-srcImage-01548 (7 tests): agc_driver_depth_clear_hle, agc_driver_depth_plane_buffer_memory, agc_driver_depth_plane_partial_store, agc_driver_depth_plane_sampling, agc_driver_depth_storage_dispatch, agc_driver_depth_surface_from_storage, agc_driver_dispatch_move_depth
VUID-vkCmdCopyImage-srcImage-01551 (7 tests): agc_driver_depth_clear_hle, agc_driver_depth_plane_buffer_memory, agc_driver_depth_plane_partial_store, agc_driver_depth_plane_sampling, agc_driver_depth_storage_dispatch, agc_driver_depth_surface_from_storage, agc_driver_dispatch_move_depth
VUID-VkPipelineColorBlendStateCreateInfo-pAttachments-00605 (1 test): agc_driver_srgb8_color_target
VUID-VkPipelineShaderStageCreateInfo-module-parameter (1 test): agc_driver_float16_unused_nan_payloads
VUID-VkPipelineShaderStageCreateInfo-pSpecializationInfo-06849 (1 test): agc_driver_float16_unused_nan_payloads
VUID-VkSamplerCreateInfo-addressModeU-01079 (1 test): agc_driver_image_gather_volume
VUID-VkSamplerCreateInfo-addressModeU-parameter (1 test): agc_driver_image_gather_volume
VUID-VkSamplerCreateInfo-addressModeV-parameter (1 test): agc_driver_image_gather_volume
VUID-VkSamplerCreateInfo-addressModeW-parameter (1 test): agc_driver_image_gather_volume

agc_driver_bda_reads_shadowed_results: skipped, the device reports watched host memory written when it imports or reads it
agc_driver_bda_reads_shadowed_results_per_build: skipped, the device reports watched host memory written when it imports or reads it
agc_driver_gcn_display_present: skipped, no display: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
agc_driver_sample_order_grid_reference: skipped, the device does not keep depth drawn at custom sample locations at those locations, even after a transition that names them: the composite of uniform renders holds the guest's depth in 0 of 3005 sloped samples, so no path can be compared with it bit for bit
agc_driver_scattered_runs: skipped, host imports are unavailable: /dev/udmabuf is not accessible
agc_driver_sparse_dma_buf_import: skipped, host imports are unavailable: /dev/udmabuf is not accessible
agc_host_import_sparse: skipped, host imports are unavailable: /dev/udmabuf is not accessible
agc_import_settle_initial_dcc: skipped, the device has no deferred settle of host imports: imports under 1 MiB wait for one only on Windows
agc_unit_shadow_completion_store: skipped, the device reports watched host memory written when it imports or reads it
guest_http_uri_locale: 
guest_net_ctl: skipped, the host has no IPv4 address
guest_video_out_color_space: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_hdr: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_high_frame_rate: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_output_mode: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_resolution_status: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_wqhd_detection: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
video_out_flip_aftergpu: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_backtoback: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_controls: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_cpuflip: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_cpuflip_in_order: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_dcc: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_deferred: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_eop: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_eop_aftergpu: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_eop_main: s

git fetch origin 8868994b056bec151255bb8248165594bd2b4d33 9ab6a281ea61c9e62a7eb07eb4cef8f19ede1c5c
git checkout 9ab6a281ea61c9e62a7eb07eb4cef8f19ede1c5c && git merge --no-ff 8868994b056bec151255bb8248165594bd2b4d33
git submodule update --init --recursive
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=ON -DANYPS5_ENABLE_SPIRV_TOOLS=ON
cmake --build build && cmake --build build --target libs
VK_DRIVER_FILES=/nonexistent/icd.json ctest --test-dir build -E '^(agc|shader|recompiler|videoout|video_out|guest_video_out|vulkan)|gpu' --timeout 120 --parallel 32 --output-on-failure
ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1 VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation \
  ctest --test-dir build -R '^(agc|shader|recompiler|videoout|video_out|guest_video_out|vulkan)|gpu' --timeout 120 --parallel 4 --output-on-failure
# failures are then re-run alone: --timeout 360 --parallel 2; new validation IDs: the test alone, twice, --parallel 1

_{Automated build and test on a contributor's NVIDIA machine; it does not replace upstream CI. A new comment is posted for each new push. Full logs.}

✅ NVIDIA GPU test: passed

PR head 3ab861c merged onto main e9247a9 · tested 2026-11-05 23:18 UTC · 25 min

agc_driver_pipeline_journal: failed in 1 of the last 10 main runs

agc_driver_image_table
guest_mmap_executable
guest_socket_options_multicast

agc_driver_submit_progress: Failed in the parallel run; also passed only on retry in 30 other runs in the last 7 days

agc_driver_linear_target_concurrent_bytes
agc_driver_linear_target_preserved_bytes
agc_driver_linear_target_publication

VUID-VkDeviceCreateInfo-pNext-pNext (593 tests): agc_compute_pipeline_replay, agc_driver, agc_driver_address_draw_space_reuse, agc_driver_address_passes, agc_driver_alias_copy_label, agc_driver_alpha_to_coverage, agc_driver_array_layer_target, agc_driver_bc_block_copy, and 585 more
VUID-VkShaderModuleCreateInfo-pCode-08737 (14 tests): agc_driver_dpp16_operands, agc_driver_dpp8_pk_fmac, agc_driver_dx10_clamp, agc_driver_f16_fma_denormal_modes, agc_driver_f16_fma_invalid_nan, agc_driver_float16_fma_rounding, agc_driver_float16_modifiers, agc_driver_float16_native, and 6 more
VUID-VkShaderModuleCreateInfo-pCode-08739 (14 tests): agc_driver_dpp16_operands, agc_driver_dpp8_pk_fmac, agc_driver_dx10_clamp, agc_driver_f16_fma_denormal_modes, agc_driver_f16_fma_invalid_nan, agc_driver_float16_fma_rounding, agc_driver_float16_modifiers, agc_driver_float16_native, and 6 more
VUID-VkShaderModuleCreateInfo-pCode-08741 (14 tests): agc_driver_dpp16_operands, agc_driver_dpp8_pk_fmac, agc_driver_dx10_clamp, agc_driver_f16_fma_denormal_modes, agc_driver_f16_fma_invalid_nan, agc_driver_float16_fma_rounding, agc_driver_float16_modifiers, agc_driver_float16_native, and 6 more
VUID-vkCmdCopyImage-srcImage-01548 (7 tests): agc_driver_depth_clear_hle, agc_driver_depth_plane_buffer_memory, agc_driver_depth_plane_partial_store, agc_driver_depth_plane_sampling, agc_driver_depth_storage_dispatch, agc_driver_depth_surface_from_storage, agc_driver_dispatch_move_depth
VUID-vkCmdCopyImage-srcImage-01551 (7 tests): agc_driver_depth_clear_hle, agc_driver_depth_plane_buffer_memory, agc_driver_depth_plane_partial_store, agc_driver_depth_plane_sampling, agc_driver_depth_storage_dispatch, agc_driver_depth_surface_from_storage, agc_driver_dispatch_move_depth
VUID-VkPipelineColorBlendStateCreateInfo-pAttachments-00605 (2 tests): agc_driver_prepared_exports, agc_driver_srgb8_color_target
VUID-VkPipelineShaderStageCreateInfo-module-parameter (1 test): agc_driver_float16_unused_nan_payloads
VUID-VkPipelineShaderStageCreateInfo-pSpecializationInfo-06849 (1 test): agc_driver_float16_unused_nan_payloads

agc_driver_bda_reads_shadowed_results: skipped, the device reports watched host memory written when it imports or reads it
agc_driver_bda_reads_shadowed_results_per_build: skipped, the device reports watched host memory written when it imports or reads it
agc_driver_gcn_display_present: skipped, no display: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
agc_driver_sample_order_grid_reference: skipped, the device does not keep depth drawn at custom sample locations at those locations, even after a transition that names them: the composite of uniform renders holds the guest's depth in 0 of 3005 sloped samples, so no path can be compared with it bit for bit
agc_driver_scattered_runs: skipped, host imports are unavailable: /dev/udmabuf is not accessible
agc_driver_sparse_dma_buf_import: skipped, host imports are unavailable: /dev/udmabuf is not accessible
agc_driver_stencil_ref_export: skipped, the device lacks VK_EXT_shader_stencil_export
agc_driver_vertex_elected_atomic_max: skipped, the device runs vertex programs in subgroups, so they keep their cross-lane reductions
agc_host_import_sparse: skipped, host imports are unavailable: /dev/udmabuf is not accessible
agc_unit_shadow_completion_store: skipped, the device reports watched host memory written when it imports or reads it
guest_http_uri_locale: 
guest_net_ctl: skipped, the host has no IPv4 address
guest_net_ctl_disconnected: skipped, the host has no IPv4 address
guest_video_out_color_space: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_device_capability: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_hdr: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_high_frame_rate: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_host_resolution_disabled: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_host_resolution_four_k: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_host_resolution_height_only: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_host_resolution_limit: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_host_resolution_malformed: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_host_resolution_width_only: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_output_mode: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_resolution_status: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
guest_video_out_wqhd_detection: skipped, no display or Vulkan device: SDL_InitSubSystem(VIDEO | GAMECONTROLLER) failed: wayland,x11 not available
video_out_flip_aftergpu: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_backtoback: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current SDL video driver (offscreen) or platform
video_out_flip_controls: skipped, no display or Vulkan device: DisplayWindow: Vulkan support is either not configured in SDL or not available in current 

git fetch origin 3ab861cd08dae4aa651627e0bacc861418710b40 e9247a966ff6cce55816ffe1d73a2679074348a4
git checkout e9247a966ff6cce55816ffe1d73a2679074348a4 && git merge --no-ff 3ab861cd08dae4aa651627e0bacc861418710b40
git submodule update --init --recursive
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=ON -DANYPS5_ENABLE_SPIRV_TOOLS=ON
cmake --build build && cmake --build build --target libs
VK_DRIVER_FILES=/nonexistent/icd.json ctest --test-dir build -E '^(agc|shader|recompiler|videoout|video_out|guest_video_out|vulkan)|gpu' --timeout 120 --parallel 32 --output-on-failure
ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1 VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation \
  ctest --test-dir build -R '^(agc|shader|recompiler|videoout|video_out|guest_video_out|vulkan)|gpu' --timeout 120 --parallel 4 --output-on-failure
# failures are then re-run alone: --timeout 360 --parallel 2; new validation IDs: the test alone, twice, --parallel 1

_{Automated build and test on a contributor's NVIDIA machine; it does not replace upstream CI. A new comment is posted for each new push. Full logs.}