[milescoviello]

What

Forspoken (PPSA03984) on main stopped once at 118 s with:

[gpu] driver failure: AGC graphics: vkAllocateMemory presentation source: Vulkan result -2
[videoout] presentation failed: AGC graphics: vkAllocateMemory presentation source: Vulkan result -2
terminate called after throwing an instance of 'std::runtime_error'

That run was on Linux, RTX 4070 Ti (12 GB), NVIDIA 615.71.09, with about 2.3 GB of the card held by other processes on the machine. The title fills the device-local heap (on that card it already reports device-local heap 0 uses 9164 of its 9164 MiB budget for textures), so when the presentation source image was recreated, its vkAllocateMemory returned VK_ERROR_OUT_OF_DEVICE_MEMORY. PresentationScaler::EnsureSourceImage asked for device-local memory only and threw, and the presentation failure terminated the title. Any card whose device-local heap the title fills can hit this.

How

• PresentationScaler::EnsureSourceImage picks the source's memory type with Context::DeviceMemoryType, as textures do: a device-local type while the heap's budget has room, else a host type. It takes the memory from the buffer pool's image path (AllocateImageMemory). On a refused device-local allocation that path reclaims cached pool memory first, then falls back to host memory through RefusedDeviceMemoryFallback.
• release() gives the memory back with ReleaseImageMemory instead of vkFreeMemory. The scaler keeps the allocation's size and memory type for that.
• Nothing else changes: the black image, the blits and the uploads are as before.

Overlap (tools/pr_overlap.py and git merge-tree of each head with this one):

• tests/PresentationPass.cpp: perf(agc): GPU descriptor heap on VK_EXT_descriptor_heap (APS5_GPU_HEAP) #9508 adds its GpuHeapEnabled() block right after residentPresentTests(device->GetContext()) in main, where this PR adds sourceImageOutsideDeviceMemoryTests(device->GetContext()). feat(agc): look up table images at run time through a device image table on the GPU heap (APS5_GPU_PATH) #7642 and feat(agc): serve lane-keyed, pointer and runtime image-table columns #10193 carry that block through perf(agc): GPU descriptor heap on VK_EXT_descriptor_heap (APS5_GPU_HEAP) #9508. feat(videoout): composite the overlay bus over the main output #10507 adds its own test call there, and feat(agc): scale the presentation with FSR 1 under ANYPS5_UPSCALE=fsr #10043 adds helpers just before main, where this PR adds its allocation-refusing wrappers. Each resolves by keeping both sides, with this PR's call above the heap block.
• PresentationScaler.cpp: feat(agc): trace device memory allocations and cache budgets #10194 changes the same two lines (the source's allocate and free) to MemoryAccounting::Allocate / Free. With feat(agc): trace device memory allocations and cache budgets #10194, BufferPool::AllocateMemory, TakeImageMemory, PutImageMemory and ReleaseImageMemory record the accounting themselves, so the merge keeps this PR's AllocateImageMemory / ReleaseImageMemory and drops feat(agc): trace device memory allocations and cache budgets #10194's two calls; proposed on feat(agc): trace device memory allocations and cache budgets #10194.

Tested

Linux (Ubuntu 26.04.1, GCC 15.2, Release, RTX 4070 Ti, NVIDIA 615.71.09) unless noted.

• Red/green: agc_driver_presentation_pass gains a case that gives the scaler a context whose vkAllocateMemory refuses every device-local memory type. It requires the source image to be created, at least one device-local refusal, and the image's memory to come from a host heap. With the test alone on main (7d24e61d28) it fails 3 of 3 with the title's message, AGC graphics: vkAllocateMemory presentation source: Vulkan result -2. With this change it passes 3 of 3, and the pool logs vkAllocateMemory presentation source: memory type 1 is out of device memory for 16384 bytes ...; placed in host memory type 0. The case skips on a device with no host memory type for the image.
• Focused tests: the 8 tests matching presentation|buffer_pool|^agc_driver_graphics$|display pass 3 of 3. On the Windows 11 VM (GT 1030, driver 582.78, WinLibs GCC 15.2.0 r7) the same 8 pass 3 of 3 on the merge base and on this head. The new case runs there too (memory type 7 refused, placed in host memory type 1).
• Khronos validation 1.4.363 over the same 8, each arm run as VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1 ctest -R "$R" -V -j4: 8 of 8 on main bd8e134 and on this head, no layer-missing line, no VUID- message on either.
• Forspoken (PPSA03984 01.020.000), warm, 16 GB cap, 250 s runs:
  • main 2e22b67 plus a local diagnostic that only records VK_EXT_device_address_binding_report events: 2 of 15 runs stopped as quoted above (at 118 s and 115 s).
  • This PR: 6 runs. 4 ran to the limit; 2 were stopped by my 16 GB host-memory cap at 218 s and 86 s (the second while a backup of the VM was writing). None stopped on the presentation source. The fallback also didn't trigger in these runs (no presentation source line), so they show nothing regressed, not the fallback in the title.

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. Merge order only: this PR goes before feat(agc): look up table images at run time through a device image table on the GPU heap (APS5_GPU_PATH) #7642 and feat(agc): scale the presentation with FSR 1 under ANYPS5_UPSCALE=fsr #10043 (agreed by EagleBearerYS) and feat(videoout): composite the overlay bus over the main output #10507 (agreed by orezraey). I proposed the same order on perf(agc): GPU descriptor heap on VK_EXT_descriptor_heap (APS5_GPU_HEAP) #9508, feat(agc): serve lane-keyed, pointer and runtime image-table columns #10193 and feat(agc): trace device memory allocations and cache budgets #10194, whose authors have not answered yet. Whichever goes second keeps both sides, as Overlap describes
• Fixes: none
• Unblocks: none (the stop above, Forspoken PPSA03984, when the device-local heap is full)
• GPUs tested: NVIDIA GeForce RTX 4070 Ti, driver 615.71.09, Ubuntu 26.04.1; NVIDIA GeForce GT 1030, driver 582.78, Windows 11 VM
• AI-assisted: yes (Claude Code)

🤖 Generated with Claude Code

[MarcZX10]

I reviewed the fallback path and test. The host-memory fallback is exercised by injecting a device-local allocation refusal, and the test verifies both the refusal and resulting heap class. I did not find an additional correctness blocker in the changed ownership path: the source allocation records its size/type and returns through ReleaseImageMemory. The hosted workflow has conventions/intake/report green; platform product jobs are skipped. No separate local Vulkan rerun was available here.

[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 9508	core/libs/prx/libSceAgcDriver/tests/PresentationPass.cpp L584

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 7642	core/libs/prx/libSceAgcDriver/tests/PresentationPass.cpp
PR 10193	core/libs/prx/libSceAgcDriver/tests/PresentationPass.cpp

[milescoviello]

@MarcZX10 thank you for the review. The Vulkan runs are the ones under Tested: the new case fails 3 of 3 on main with the title's exact message and passes 3 of 3 here, the 8 tests run clean under VK_LAYER_KHRONOS_validation on both sides (no new VUID), and the Windows VM (GT 1030) passes 8 of 8, 3 times, where the new case refuses memory type 7 and the source lands in host memory type 1.

AI-assisted: yes (Claude Code)

[MarcZX10]

One overlap follow-up after manually checking the bot’s full report: #7642 and #10193 are also open and both add the same GpuHeapEnabled() / OpenVulkanTestDevice() block after residentPresentTests(device->GetContext()) in PresentationPass.cpp::main. The description lists #9508, #10043 and #10507 but omits these two. The edits can be retained together with this PR’s new fallback test, but please add both to the overlap note so the merge-time resolution is explicit.

[milescoviello]

@MarcZX10 thanks for checking the full report. Confirmed: #7642 and #10193 both add that GpuHeapEnabled() / OpenVulkanTestDevice() block right after residentPresentTests(device->GetContext()), which is where this PR adds sourceImageOutsideDeviceMemoryTests(device->GetContext()). The tool missed #10193 because it conflicts with main. I've added both to the overlap note: keeping this PR's call and their block resolves either order.

AI-assisted: yes (Claude Code)

[MarcZX10]

Thanks for confirming that #7642 and #10193 add the same GpuHeapEnabled() / OpenVulkanTestDevice() fixture block. Both PRs are still open, while this PR checklist remains Depends on: none. CONTRIBUTING.md requires the authors to agree which overlapping PR goes first and declare it with Depends on. Please coordinate with #7642 and #10193 and update the checklist (and have the other affected checklist(s) reflect the agreed order). I have not modified or rebased any branch. The hosted policy checks currently pass; product jobs are skipped.

[milescoviello]

@MarcZX10 I've asked both authors, on #7642 and #10193. I proposed this PR goes first: #7642 is a draft with changes requested, #10193 depends on #7642 and needs a rebase, and it's a merge order with no code dependency either way. The checklist now records that proposal and that they haven't answered yet. I'll update it, and rebase if they'd rather go first, once they reply.

AI-assisted: yes (Claude Code)

[milescoviello]

Thanks for merging, @oneandonlydean, and @MarcZX10 for the review.

AI-assisted: yes (Claude Code)