[GSD-12441] DG2 (Arc A770) not detected by compute-runtime when using xe kernel driver

[markussiebert]

Description

OpenCL and Level Zero report 0 platforms/drivers when Intel Arc A770 (DG2, device ID 56a0) is using the xe kernel driver. The same hardware works correctly with the i915 driver.

This appears related to #853 (Arc Pro A40 with xe), which was reported as fixed in 26.01.36711.4, but the issue persists for DG2/Arc A770 on the latest release.

Environment

• GPU: Intel Arc A770 (DG2, PCI ID 8086:56a0, rev 08)
• OS: Proxmox VE 9 (Debian Trixie)
• Kernel: 6.17.13-1-pve
• Kernel driver: xe (via xe.force_probe=56a0)
• compute-runtime: 26.05.37020.3-0
• IGC: 2.28.4+20760
• libigdgmm: 22.9.0

Reproduction

1. Bind Arc A770 to xe driver:

   modprobe xe force_probe=56a0
   echo "0000:0a:00.0" > /sys/bus/pci/drivers/xe/bind
   

2. Run clinfo:

   Number of platforms                               0
   

3. Level Zero also fails:

   # zeInit returns 0x78000001 (ZE_RESULT_ERROR_DEPENDENCY_UNAVAILABLE)
   # zeDriverGet returns 0 drivers

strace shows successful device communication

The runtime does discover and query the device via xe ioctls — all calls succeed:

openat(AT_FDCWD, "/dev/dri/by-path/pci-0000:0a:00.0-render", O_RDWR|O_CLOEXEC) = 4
ioctl(4, DRM_IOCTL_VERSION, ...) = 0
ioctl(4, DRM_IOCTL_XE_DEVICE_QUERY, ...) = 0   # multiple successful queries

But then the runtime silently drops the device — no platform is registered.

dmesg (xe driver)

xe 0000:0a:00.0: [drm] Found dg2/g10 (device ID 56a0) discrete display version 13.00 stepping C0
xe 0000:0a:00.0: [drm] VRAM[0]: Actual physical size 0x0000000400000000, usable size exclude stolen 0x00000003fa000000, CPU accessible size 0x0000000010000000

Expected behavior

compute-runtime should support DG2/Arc via the xe driver. The motivation for using xe over i915 is proper runtime power management — xe achieves suspended state (~1W idle headless), while i915 reports unsupported for runtime PM and keeps the GPU active at all times.

Workaround

Switching back to i915 makes compute work, but the GPU cannot enter deep sleep.

Questions

1. Is DG2 compute via xe driver intended to be supported?
2. Is there a timeline for xe+DG2 compute support?

[PopoviciGabriel]

I tried to make this issue happen on my computer. I saw something similar.

Environment

• OS: Ubuntu 24.04

• Kernel: 6.8

• GPU: Intel

• intel-compute-runtime: the one from the repository

• IGC: the latest version I could find

Steps I followed

1. I installed the compute runtime from the repository

2. I checked the platform with clinfo

3. I ran a program to test the kernel

After running for a little while the program seems to get stuck when it is waiting for everything to finish like around clFinish or when the queue is done.

From what I can see the problem might be with how the command queue's synchronized or how events are handled inside the runtime. I am not really sure yet. It could have something to do with how the command submission waits for the GPU to finish.

If it helps I can also give you:

• the clinfo output

• debug logs using NEOReadDebugKeys=1

• the stack trace, from the process that got stuck

Let me know if there are logs or traces that would help figure out what is going on with this issue.

[markussiebert]

Update: Resolved by enabling Resizable BAR in BIOS.

The root cause was not the xe driver itself, but the lack of Resizable BAR. With only 256MB BAR, the compute-runtime silently rejected the DG2 device when using the xe driver (while i915 worked fine with small BAR).

After enabling Above 4G Decoding + Re-Size BAR Support in BIOS:

• BAR changed from [size=256M] to [size=16G]
• clinfo now reports 1 platform with full OpenCL 3.0 support via xe driver

Suggestion: The compute-runtime should log a more informative error when rejecting a device due to small BAR, rather than silently returning 0 platforms.

[kgibala]

Hi @markussiebert,

Thank you for your contribution.

Default setup with forced xeKMD module on DG2:

xe.force_probe=56a0 i915.force_probe=!56a0 modprobe.blacklist=i915 modules-load=xe 

lspci -vvv -k -s 0000:03:00.0
03:00.0 VGA compatible controller: Intel Corporation DG2 [Arc A770] (rev 08) (prog-if 00 [VGA controller])
        Subsystem: Intel Corporation DG2 [Arc A770]
        Control: I/O+ Mem+ BusMaster+ SpecCycle- MemWINV- VGASnoop- ParErr- Stepping- SERR- FastB2B- DisINTx-
        Status: Cap+ 66MHz- UDF- FastB2B- ParErr- DEVSEL=fast >TAbort- <TAbort- <MAbort- >SERR- <PERR- INTx-
        Latency: 0, Cache Line Size: 64 bytes
        Interrupt: pin ? routed to IRQ 173
        IOMMU group: 18
        Region 0: Memory at 51000000 (64-bit, non-prefetchable) [size=16M]
        Region 2: Memory at 6000000000 (64-bit, prefetchable) [size=16G]
        Expansion ROM at 52000000 [disabled] [size=2M]
        Capabilities: <access denied>
        Kernel driver in use: xe
        Kernel modules: i915, xe

clinfo -l
Platform #0: Intel(R) OpenCL Graphics
 `-- Device #0: Intel(R) Arc(TM) A770 Graphics

clinfo | grep --color -E "Device Name|Device Vendor ID|Driver Version|Max compute units|Global memory size|Max memory allocation" | awk '!x[$0]++'
  Device Name                                     Intel(R) Arc(TM) A770 Graphics
  Device Vendor ID                                0x8086
  Driver Version                                  26.05.37020.3
  Max compute units                               512
  Global memory size                              16225243136 (15.11GiB)
  Max memory allocation                           4294959104 (4GiB)

Case with: xe.vram_bar_size=256

xe.force_probe=56a0 i915.force_probe=!56a0 modprobe.blacklist=i915 modules-load=xe xe.vram_bar_size=256

sudo dmesg | grep bar
...
[    3.178339] xe 0000:03:00.0: [drm] Attempting to resize bar from 16384MiB -> 256MiB

clinfo -l
WARNING: Resizable BAR not detected for device 0000:03:00.0

clinfo
WARNING: Resizable BAR not detected for device 0000:03:00.0
Number of platforms                               0

ICD loader properties
  ICD loader Name                                 OpenCL ICD Loader
  ICD loader Vendor                               OCL Icd free software
  ICD loader Version                              2.2.12
  ICD loader Profile                              OpenCL 3.0

As you can see proper warning should be visible. Limitation is described in our FAQ:

• https://github.com/intel/compute-runtime/blob/master/FAQ.md#why-am-i-seeing-resizable-bar-not-detected-warning-and-the-gpu-driver-is-not-working

We are aware of this and have noted it as a feature request for further evaluation:

• [GSD-12001] OpenCL on Arc B580 (Battlemage) shows “Number of platforms: 0” on Ubuntu 24.04 (xe) #872