STAR WARS Zero Company™

Community report notes

The game boots, loads its bundled modules, compiles ~200 compute pipelines and presents black frames. After ~50–100 s the guest stops flipping (flipArg stops at ~1900–2600) and the window stays black at 0 fps for good. No logo is ever shown.

NVIDIA Xid 109 on every run (one per run, from dmesg):

NVRM: Xid (PCI:0000:01:00): 109, pid=763199, name=kyty_emulator, channel 0x0000001f, errorString CTX SWITCH TIMEOUT, Info 0x1bc028
NVRM: Xid (PCI:0000:01:00): 109, pid=766631, name=kyty_emulator, channel 0x0000001f, errorString CTX SWITCH TIMEOUT, Info 0x1bc028
NVRM: Xid (PCI:0000:01:00): 109, pid=768524, name=kyty_emulator, channel 0x0000001f, errorString CTX SWITCH TIMEOUT, Info 0xbc028
NVRM: Xid (PCI:0000:01:00): 109, pid=770285, name=kyty_emulator, channel 0x0000001f, errorString CTX SWITCH TIMEOUT, Info 0x1bc028
NVRM: Xid (PCI:0000:01:00): 109, pid=771820, name=kyty_emulator, channel 0x0000001f, errorString CTX SWITCH TIMEOUT, Info 0xbc028

So a shader in one batch never finishes, the driver times the context out, and that batch's timeline tick is never signaled. Afterwards the GPU is idle (~28 W) and the device is not reported lost, so MasterSemaphore::Wait (waitSemaphores(..., UINT64_MAX)) waits forever.

Where the emulator waits (59183db, stripped Release binary symbolized with its lld map). Guest GPU thread:

MasterSemaphore::Wait
CommandScheduler::Wait
BufferCache::DownloadBufferMemory<false>
BufferCache::ReadMemory
TextureCache::MaterializeColorClear          (another run: RenderContext::HandleFault <- CommandProcessor::DrawIndirectMulti)
TextureCache::FindImage
RenderExecutor::ResolveRenderColorTarget
RenderExecutor::PrepareDrawRenderState / DrawIndex
CommandProcessor::DrawIndexOffset / ProcessPm4 / Process
GuestGpu::Process / GuestGpu::ThreadRun

The game's RHISubmissionThread then waits in Gen5::AgcSuspendPoint -> GuestGpu::SuspendPoint (the previous suspend point is never released), which stops the frame loop.

With a local, logging-only change to MasterSemaphore::Wait (5 s timeout slices; the hang is the same without it):

MasterSemaphore::Wait stalled 15 s: waiting for tick 14622, GPU reached 14621, recorded up to 14623

The batch the GPU thread has just submitted never completes. No wait semaphores are attached to it. --present-mode Immediate does not change anything.

Vulkan validation (--vulkan-validation true): no validation errors before the hang (they are fatal in this build; the run reached the same hang). Repeated warnings:

vkCmdDrawIndexed(): Inside [VK_SHADER_STAGE_FRAGMENT_BIT] [Output variable, Location 0, "out_mrt_0"] with a numeric type of FLOAT but VkRenderingInfo::pColorAttachments[0].imageView is created with VK_FORMAT_R32_UINT (numeric type of UINT)
(same for VK_FORMAT_R32G32B32A32_UINT)

Fragment exports to integer render targets are emitted as float. UE5 uses R32_UINT / RGBA32_UINT targets (e.g. the Nanite visibility buffer); undefined values there could send a later GPU-driven pass into a loop that never ends, which would match the Xid 109. This is a guess, not proven.

On 2f4b8a3 and U59 the game also calls three unresolved Agc stubs (uZW-mqsxkrM AgcCbBranchGetSize, 7Wa3aeJgeVU AgcBranchPatchSetThenTarget, GBCh3zCihoU AgcDcbSetCxRegistersIndirectGetSize). On 59183db these are implemented and no Agc stub is called, but the hang is the same. The only stubs called on 59183db are NpCppWebApi / Http2 / Net / Ssl (offline).

Steps to reproduce

1. Use a dump of PPSA18054 01.000.003 (the folder with eboot.bin, sce_sys, sce_module, prx).
2. Run: kyty_emulator --game <PPSA18054 folder> --gpu 0 --screen-width 1920 --screen-height 1080 --amd-cpu --shader-validation false --skip-notice-screen true --printf-direction Console
3. Wait about 1–2 minutes: the window stays black, the guest stops flipping, and dmesg shows Xid 109 for kyty_emulator.

Expected behavior

The game shows its startup logos and reaches the main menu.

Last working build / first broken build

(no build has booted it)