[GSD-13011] Arc B580 (Battlemage G21, 8086:b080): Level Zero kernel compilation fails with "could not create an engine" despite IGC 2.36.3 + SYCL having BMG definitions

[GIS-blackCaat]

Pre-submission Checklist

• I am using the latest GPU driver version (releases)
• I have searched for similar issues and found none

---

System Information

GPU Hardware

Intel Arc B580 (Battlemage G21) [8086:b080]

DRI Devices Information

$ ls -la /dev/dri/
total 0
drwxr-xr-x   3 root root        100  7月  5 22:05 .
drwxr-xr-x  22 root root       4940  7月  5 22:06 ..
drwxr-xr-x   2 root root         80  7月  5 22:05 by-path
crw-rw----+  1 root video  226,   1  7月  5 22:05 card1
crw-rw----+  1 root render 226, 128  7月  5 22:05 renderD128

$ ls -la /dev/dri/by-path/
total 0
drwxr-xr-x 2 root root  80  7月  5 22:05 .
drwxr-xr-x 3 root root 100  7月  5 22:05 ..
lrwxrwxrwx 1 root root   8  7月  5 22:05 pci-0000:00:02.0-card -> ../card1
lrwxrwxrwx 1 root root  13  7月  5 22:05 pci-0000:00:02.0-render -> ../renderD128

GPU Detailed Information (lspci output)

$ lspci -nn -k -s 0000:00:02.0
00:02.0 VGA compatible controller [0300]: Intel Corporation Device [8086:b080] (rev 04)
        Subsystem: Xiaomi Device [1d72:2424]
        Kernel driver in use: xe
        Kernel modules: xe

Driver Version

26.22.38646.4-0

Installed GPU Driver Packages

$ dpkg --list | grep -iE "igc|gmm|opencl|level-zero|fc|ocloc|libze|intel-oneapi"

ii  clinfo                                         3.0.23.01.25-1build1
ii  intel-igc-cm                                   1.0.176.54074-1029~24.04
ii  intel-igc-core-2                               2.36.3
ii  intel-ocloc                                    26.22.38646.4-0
ii  intel-oneapi-runtime-dpcpp-sycl-opencl-cpu     2026.1.0-235
ii  intel-oneapi-runtime-opencl                    2026.1.0-235
rc  intel-opencl-icd                               26.22.38646.4-0
ii  libigc1                                        1.0.17791.16-1032~24.04
ii  libigdfcl1                                     1.0.17791.16-1032~24.04
ii  libigdgmm12:amd64                              22.10.0
ii  libze-intel-gpu1                               26.22.38646.4-0
ii  libze1                                         1.17.44.0-1022~24.04
ii  ocl-icd-libopencl1:amd64                       2.3.2-1build1

Driver Installation Details

• Installation method: Intel official APT repository + oneAPI installer
• Added Intel repository for GPU driver packages
• Installed oneAPI 2026.1.0 base toolkit
• PyTorch XPU installed via pip from Intel channel
• No custom kernel parameters for GPU

Linux Distribution

Ubuntu 24.04 LTS

Kernel Version & Boot Parameters

$ uname -r
6.17.0-35-generic

$ cat /proc/cmdline
BOOT_IMAGE=/boot/vmlinuz-6.17.0-35-generic root=UUID=c9dfbc53-0f51-464e-a180-fe8dc56e6061 ro quiet splash i8042.nopnp=1 i8042.dumbkbd=1 vt.handoff=7

$ lsmod | grep -E 'i915|xe'
xe                   3874816  10
gpu_sched              65536  1 xe
drm_gpuvm              49152  1 xe
drm_gpusvm_helper      36864  1 xe
drm_buddy              28672  1 xe
drm_ttm_helper         16384  1 xe
ttm                   126976  2 drm_ttm_helper,xe
drm_exec               12288  2 drm_gpuvm,xe
drm_suballoc_helper    24576  1 xe
drm_display_helper    290816  1 xe
cec                    98304  2 drm_display_helper,xe
i2c_algo_bit           16384  1 xe
intel_vsec             24576  3 intel_pmc_ssram_telemetry,pmt_telemetry,xe
video                  77824  1 xe

---

Issue Description

Actual Behavior

GPU is detected and basic operations work, but kernel compilation fails for any non-trivial computation. Specifically:

1. torch.xpu.is_available() returns True
2. torch.xpu.device_count() returns 1
3. Data transfer to XPU device works (e.g., torch.randn(10,10).to("xpu"))
4. Element-wise operations work (e.g., x + 1.0)
5. Matrix multiplication fails: x @ x.T → RuntimeError: could not create an engine
6. torch.xpu.get_device_name(0) segfaults

The SYCL runtime has Battlemage device definitions (intel_gpu_bmg_g21) and IGC has Xe2 code (AllowSIMD16DropForXE2Plus), suggesting the PCI ID → architecture mapping or kernel compilation flag path is broken.

Expected Behavior

Kernel compilation should succeed for all operations on the Arc B580 (Battlemage G21), including matrix multiplication and other non-trivial GPU computations. torch.xpu.get_device_name(0) should return the device name without segfaulting.

Reproduction Rate

Always reproduces - 100%

Steps to Reproduce

1. Install Intel GPU driver 26.22.38646.4-0 with IGC 2.36.3
2. Install PyTorch 2.6.0+xpu from Intel channel
3. Run the following Python script:

import torch
import intel_extension_for_pytorch

print("XPU available:", torch.xpu.is_available())
print("Device count:", torch.xpu.device_count())

x = torch.randn(10, 10).to("xpu")
y = x + 1.0  # works
print("Element-wise add OK")

# This fails:
y = x @ x.T
print("Matmul OK")

4. Observe RuntimeError: could not create an engine on the matmul line
5. Optionally, run torch.xpu.get_device_name(0) to observe segfault

---

Regression Information

Is this a regression?

• Yes, this is a regression - functionality that previously worked is now broken

This is a new GPU (Battlemage G21) and has never worked correctly with this software stack. Not a regression per se, but a missing/incomplete platform enablement.

---

Diagnostics & Logs

API Call Logs

Not yet collected with Unitrace. Can provide if needed.

strace Logs

Not yet collected. Can provide if needed.

System Logs / dmesg Output

$ dmesg | grep -i -E 'i915|xe|drm|gpu|bmg|battlemage|b080|error|fault|fail'
(No relevant GPU error messages found in dmesg)

Backtrace

Not yet collected. The get_device_name(0) call segfaults - can provide gdb backtrace if needed.

---

Reproducer Information

Source Code / Reproducer

import torch
import intel_extension_for_pytorch

x = torch.randn(10, 10).to("xpu")
y = x @ x.T  # RuntimeError: could not create an engine

Command Line / Application Details

Run the above Python script with PyTorch 2.6.0+xpu and Intel oneAPI 2026.1.0 environment sourced.

---

Additional Information

oneAPI Version

oneAPI DPC++ 2026.1.0-235

Additional Notes

• Investigation performed:
  • strings libsycl.so.8 | grep bmg → intel_gpu_bmg_g21, intel_gpu_bmg_g31 — SYCL knows about BMG
  • strings /usr/local/lib/libigc.so.2 | grep Xe2 → AllowSIMD16DropForXE2Plus, LscInstructionXe2 — IGC has Xe2 code
  • oneAPI 2026.1 device_architecture.def lists BMG G21 and G31
• Combinations tried:
  • torch 2.6.0+pip + SYCL 2025.0.2 → this error (could not create an engine)
  • torch 2.6.0+conda + SYCL 2025.3.2 → "device architecture not supported"
  • torch 2.8.0+oneAPI + SYCL 9 → ABI mismatch
• Suspected root cause: The Level Zero → IGC kernel compilation path may not correctly map the Battlemage G21 PCI ID (8086:b080) to the Xe2 compilation target, despite both SYCL and IGC having the necessary definitions.
• Related to prior issue Arc B580 (Battlemage G21, 8086:b080): Level Zero kernel compilation fails despite IGC 2.36.3 + SYCL having BMG definitions #947 (submitted without template)

Please investigate whether the kernel compilation path correctly handles the Battlemage G21 PCI ID. Happy to provide any additional logs or run further diagnostics.

[kgibala]

Hi @GIS-blackCaat,

Thank you for your contribution and for bringing this to our attention.

The issue is currently being reviewed and addressed through our internal triage process.

We will provide an update here as soon as there is meaningful progress to share.

Thank you for your patience and understanding.