Skip to content

Releases

v0.6

Choose a tag to compare

@osy osy released this 17 Nov 17:08

New

  • Vulkan (Venus) works out of the box: with the driver installed, Vulkan apps now find the "Virtio-GPU Venus" device without any VK_DRIVER_FILES override. Before, the Vulkan driver failed to load from the driver store, and failed to start its command ring ("failed to allocate/map ring shmem") on the WDDM 2 kernel driver (#10, utmapp/mesa#8)
  • x64 Vulkan apps on Windows on Arm: the Vulkan driver is now packaged as ARM64X, so x64-emulated and ARM64EC Vulkan games and apps (most Steam games, including Quake II RTX) can use Venus, not only native ARM64 apps (#11, utmapp/mesa#11)
  • 10-bit display modes: the display now offers each resolution in 10-bit (R10G10B10A2) too, and 10-bit fullscreen output is shown at full depth instead of being treated as 8-bit. Games that list resolutions in 10-bit no longer show an empty list; FORSPOKEN's Display Settings showed "0 x 0" and its resolution and screen mode couldn't be changed. Requires UTM 5.0.9 or later; on older UTM versions 10-bit fullscreen output shows wrong colours (#22, utmapp/mesa#16)

Fixes

  • Smoother frame pacing: vsync now ticks at the display's exact refresh rate. It previously ran 4–15% slow, so 5–11% of frames presented with vsync were never shown (#19)
  • Boot screen: the Windows boot spinner now stays on screen until the Welcome screen, instead of showing "Display output is not active." and a black frame while Windows starts (#16, #21)
  • Games that recreate swap chains or shared textures often: no longer lose the GPU (DXGI_ERROR_DEVICE_RESET) after a while, e.g. after 10–12 seconds in FORSPOKEN's Display Settings. Released shared resources are now freed right away instead of staying allocated until the game exits (utmapp/mesa#15)
  • Windows that are opened and closed over a long session: the GPU memory behind a closed window's swap chain is now released, instead of being held by the desktop compositor (DWM) until it restarts. Over time that could exhaust GPU memory for every app (#20, utmapp/mesa#14)
  • Games and apps with strict Control Flow Guard: the D3D driver can now be loaded into processes that enable strict CFG. Some Unreal Engine D3D11 games stopped at "A D3D11-compatible GPU (Feature Level 11.0, Shader Model 5.0) is required to run the engine" (utmapp/mesa#13)
  • Quake II RTX and other Vulkan apps that allocate past a descriptor pool's declared sizes: no longer crash when a map loads (utmapp/mesa#10)
  • D3D and Vulkan from Windows services and the guest agent (session 0): devices can now be created there; before, no GPU adapter was found (utmapp/mesa#9)

v0.5

Choose a tag to compare

@osy osy released this 29 Oct 18:38

Fixes

  • Windows 11 desktop: window drop shadows now render on both D3DMetal and DXMT. DWM builds its shadow cut-out mask with target-independent rasterization, and the pixel shader's coverage input now reports full coverage there (utmapp/UTM#7993)
  • D3D11 games and apps that create a render-target view every frame: frame rate no longer drops steadily over time and then stalls for several seconds. The driver now lets the D3D11 runtime destroy released objects after each present (utmapp/UTM#7965)
  • Subnautica 2 and other Unreal Engine 5 games: no longer quit at launch with "RHI D3D12 is not supported on your system". The D3D12 driver now reports the host's 64-bit atomic capabilities, which UE5's SM6 renderer requires
  • 32-bit (x86) games and apps, including all 32-bit D3D9 titles: no longer lose the GPU and crash a few seconds after launch. Pointer-sized values and D3D12 descriptor handles are now encoded correctly in the 32-bit driver
  • D3D9 and D3D12 games and apps using antialiasing (MSAA): no longer lose the GPU, and windowed D3D9 MSAA output is no longer blank. Multisampled render targets are now created and resolved correctly
  • D3D9 games and apps: fixed random crashes at device creation, caused by a background driver thread that outlived the driver DLL being unloaded
  • Display resolution: resizing the VM window or entering full screen now changes the Windows resolution to match (utmapp/UTM#8000, #12)
  • Display scaling: your chosen resolution and scaling now persist across window resizes, guest restarts and UTM relaunches. Previously Windows could switch to 125% after a resize, or come back at 1024x768 and 100%

v0.4

Choose a tag to compare

@osy osy released this 15 Oct 16:54
viogpu3d: hold a reference across the RES_INFO and RES_RELEASE escapes

Both escapes resolved their cookie or handle to a bare allocation
pointer and dropped m_PendingCreateLock before using it.  Their payload
is unvalidated user input, and a DxgkDdiDestroyAllocation in another
process could free the allocation between the lookup and IsOpenOn /
EscapeResourceInfo.  Resolve through EscapeAllocationRef, which takes a
reference under the lock the way AllocationByResIdRef does, and release
it when the escape is done.

Assisted-by: Claude:claude-opus-5-5

v0.3: viogpu3d: back RES_INFO blob windows with a synchronous host map

Choose a tag to compare

@osy osy released this 01 Sep 06:19
The RES_INFO placement escape handed back a window that was placed and
process-mapped but not yet host-backed: the MAP_BLOB reached the host
only when the UMD's follow-up DMA packet ran, and the UMD's only way to
learn that had happened was a full scheduler drain behind every DMA
packet already queued on the context (~12 ms under load, on every
mappable blob create).

Add CtrlQueue::ResourceMapBlobSync -- the ResourceUnmapBlobSync wait-ctx
pattern applied to MAP_BLOB, with the host response checked the way
CreateResourceBlob checks it -- and issue it from EscapeResourceInfo
once the blob is created and unmapped. RES_INFO now returns a window
that is placed, process-mapped, and host-backed in one answer, and
because the control queue is FIFO the acknowledgement also proves the
blob's CREATE_BLOB was processed, so a ring command may name the res_id
immediately with no ordering roundtrip.

A host that answers but refuses the map fails the escape with
STATUS_DEVICE_NOT_READY instead of handing out an unbacked window. The
blt-present path passes no mapping device: it only reads the descriptor
and must not establish a mapping the UMD never asked for.

Assisted-by: Claude:claude-fable-5

v0.2.1

v0.2.1 Pre-release
Pre-release

Choose a tag to compare

@osy osy released this 02 Aug 08:39
viogpuap: skip SetDisplayConfig when the resolution is unchanged

The resolution event fires for every host display event, not only for
actual mode changes, and SetDisplayConfig with SDC_APPLY forces a
modeset -- a visible blank plus a DWM swapchain rebuild -- even when
the supplied mode equals the current one. Every no-op monitors config
from the client therefore blanked the guest display, and guest-initiated
resolution changes were fought by the reapply.

Compare GetCustomResolution against GetCurrentResolution and skip the
apply (with a log line) when they already match.

That comparison also makes SetResolution the branch taken when the
display config could not be read at all, so bounds-check what it
indexes: m_Index against the path array, and the path's modeInfoIdx
against the mode info array. An inactive path carries
DISPLAYCONFIG_PATH_MODE_IDX_INVALID and a failed QueryDisplayConfig
leaves the arrays NULL, either of which would have been written out of
bounds.

Assisted-by: Claude:claude-fable-5
Assisted-by: Claude:claude-opus-5[1m]

v0.2

v0.2 Pre-release
Pre-release

Choose a tag to compare

@osy osy released this 29 Jul 16:32
viogpu: stop reading an IO_STATUS_BLOCK the I/O manager never wrote

Installing the driver over itself bugchecked 0x50 PAGE_FAULT_IN_NONPAGED_AREA
in InjectFramebufferResource, reached from the IRP_MJ_PNP hook while PnP was
restarting the device.

The hook queried the PDO's boot configuration with
IoBuildSynchronousFsdRequest() and then read the resource list out of the
IO_STATUS_BLOCK it had passed in.  That helper reports its result by writing
the caller's status block from the I/O manager's special kernel APC, and the
hook runs inside PnP enumeration, where that APC can be deferred past the
hook's own return.  The status block therefore still held whatever the stack
happened to contain -- at the crash, IoCallDriver had returned
STATUS_NOT_SUPPORTED (pci has no boot config to report once the device is
being restarted) while Information held a stale stack address, which the
resource walk then followed until it ran off the top of the kernel stack.
Nothing checked the IoCallDriver status or the list pointer, and the IRP was
leaked whenever the APC did not run.

  - The query IRP is now allocated and freed here, with a completion routine
    that signals an event and returns STATUS_MORE_PROCESSING_REQUIRED, so the
    result is read straight out of Irp->IoStatus with no dependency on APC
    delivery and no ambiguity about who frees the IRP.
  - A failed query, or one that yields no list, forwards the original IRP to
    the original dispatch, so a device with no boot configuration behaves
    exactly as it would with the dispatch unpatched.  A failure to inject no
    longer fails the query either: the unmodified list is passed up instead.
  - The fixup only applies to the PDO the fixup was installed for.  The patched
    dispatch belongs to \Driver\pci and is shared by every PCI device, so
    IRP_MN_QUERY_RESOURCES for unrelated devices was being answered with a
    framebuffer resource injected into their boot configurations.
  - The dispatch is restored only while it is still ours, since a driver that
    patched on top cannot be unlinked by putting the saved pointer back.

Validated: three install/verify cycles from a cold boot, device Started each
time, D3D11 device creation on the VirtIO adapter at all five feature levels,
no bugcheck; desktop renders after a clean cycle.

Assisted-by: Claude:claude-opus-5

v0.1

v0.1 Pre-release
Pre-release

Choose a tag to compare

@osy osy released this 24 Jul 03:17
viogpu3d: vsync KTIMER tick, lookaside fence contexts, flip-health co…