different `libsycl.so.8.0.0` different `UR_DEVICE_INFO_TYPE` for device

[drzraf]

Describe the bug

sycl-ls brings inconsistent results (using different release of libsycl.so.8). It fails to recognize my HD Graphics 620 capabilities.

In some cases it detects the device as a SYCL usable GPU but in some others, it doesn't.

In all of the below attempts, the sycl-ls binary tested below comes from https://github.com/intel/llvm/releases/download/nightly-2024-06-03/sycl_linux.tar.gz

Testing libsycl.so.8.0.0-0 from nightly-2024-06-03

$ ls -lh libsycl.so.8.0.0-0
5,4M june   3  2024 libsycl.so.8.0.0-0

$ LD_LIBRARY_PATH=/tmp/nightly-2024-06-03... sycl-ls
[opencl:cpu][opencl:0] Portable Computing Language, cpu-haswell-Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz OpenCL 3.0 PoCL HSTR: cpu-x86_64-pc-linux-gnu-haswell [5.0+debian]
[opencl:cpu][opencl:1] Intel(R) OpenCL, Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz OpenCL 3.0 (Build 0) [2025.20.6.0.04_224945]
[cuda:gpu][cuda:0] NVIDIA CUDA BACKEND, NVIDIA GeForce 940MX 5.0 [CUDA 12.2]

$ strace -e trace=openat sycl-ls 2>&1|grep libsycl.so|grep -v ENOENT
openat(AT_FDCWD, "/nightly-2024-06-03.../libsycl.so.8", O_RDONLY|O_CLOEXEC) = 3

Result:

The iGPU is wrongly seen as a CPU. The additional GeForce is found.

Using 2025 intel-oneapi-* .deb packages

Installing intel-oneapi-common-var and intel-oneapi-compiler-dpcpp-cpp-runtime-2025.2 (among 5+GB of packages)

$ ls -lh /opt/intel/oneapi/compiler/2025.2/lib/
12M juin   4 23:55 /opt/intel/oneapi/compiler/2025.2/lib/libsycl.so.8.0.0

(This newer libsycl is twice the size)

$  . /opt/intel/oneapi/setvars.sh
$ sycl-ls
[opencl:gpu][opencl:0] Intel(R) OpenCL, Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz OpenCL 3.0 (Build 0) [2025.20.6.0.04_224945]

strace excerpt.

(openat(AT_FDCWD, "/opt/intel/oneapi/compiler/2025.2/lib/libsycl.so.8", O_RDONLY|O_CLOEXEC) = 3)

Result

The Intel HD Graphics 620 is correctly identified as a GPU (no trace of GeForce). It's better, but still a bit weird

But it's more tricky than that.

I went on and compiled some code using Platform.get_backend() modeled after original source code :
$ /opt/intel/oneapi/compiler/2025.2/bin/icpx -I/opt/intel/oneapi/compiler/2025.2/include -fsycl -ggdb -O0 test.cpp

and this are the results:

using nightly-2024-06-03's libsycl.so.8

[opencl:unknown][opencl:0] Intel(R) OpenCL, Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz OpenCL 3.0 (Build 0) [2025.20.6.0.04_224945]
[opencl:unknown][opencl:1] Portable Computing Language, cpu-haswell-Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz OpenCL 3.0 PoCL HSTR: cpu-x86_64-pc-linux-gnu-haswell [5.0+debian]

(no GeForce, "unknown" Intel devices)

and using the regular 2025.2 library:

[opencl:cpu][opencl:0] Intel(R) OpenCL, Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz OpenCL 3.0 (Build 0) [2025.20.6.0.04_224945]

(HD Graphics 620 found as CPU)

What's going on ?

To reproduce

Environment

• OS: Ubuntu 24.04
• Target device and vendor: HD Graphics 620
• DPC++ version: Intel(R) oneAPI DPC++/C++ Compiler 2025.2.0 (2025.2.0.20250605)
• sycl-ls --verbose

Platforms: 3
Platform [#1]:
    Version  : OpenCL 3.0 LINUX
    Name     : Intel(R) OpenCL
    Vendor   : Intel(R) Corporation
    Devices  : 1
        Device [#0]:
        Type       : cpu
        Version    : OpenCL 3.0 (Build 0)
        Name       : Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz
        Vendor     : Intel(R) Corporation
        Driver     : 2025.20.6.0.04_224945
        Aspects    : cpu fp16 fp64 online_compiler online_linker queue_profiling usm_device_allocations usm_host_allocations usm_shared_allocations usm_system_allocations usm_atomic_host_allocations usm_atomic_shared_allocations atomic64 ext_intel_device_info_uuid ext_oneapi_srgb ext_oneapi_native_assert ext_intel_legacy_image ext_oneapi_ballot_group ext_oneapi_fixed_size_group ext_oneapi_opportunistic_group ext_oneapi_tangle_group ext_oneapi_limited_graph ext_oneapi_private_alloca
        info::device::sub_group_sizes: 4 8 16 32 64


Platform [#2]:
    Version  : OpenCL 3.0 PoCL 5.0+debian  Linux, None+Asserts, RELOC, SPIR, LLVM 16.0.6, SLEEF, DISTRO, POCL_DEBUG
    Name     : Portable Computing Language
    Vendor   : The pocl project
    Devices  : 1
        Device [#1]:
        Type       : cpu
        Version    : OpenCL 3.0 PoCL HSTR: cpu-x86_64-pc-linux-gnu-haswell
        Name       : cpu-haswell-Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz
        Vendor     : GenuineIntel
        Driver     : 5.0+debian
        Aspects    : cpu fp64 online_compiler online_linker queue_profiling usm_device_allocations usm_host_allocations usm_shared_allocations atomic64 ext_oneapi_srgb ext_intel_legacy_image ext_oneapi_ballot_group ext_oneapi_fixed_size_group ext_oneapi_opportunistic_group ext_oneapi_tangle_group ext_oneapi_limited_graph ext_oneapi_private_alloca
        info::device::sub_group_sizes: 1 2 4 8 16 32 64 128 256 512

# unrelated
Platform [#3]:
    Version  : CUDA 12.2
    Name     : NVIDIA CUDA BACKEND
    Vendor   : NVIDIA Corporation
    Devices  : 1
        Device [#0]:
        Type       : gpu
        Version    : 5.0
        Name       : NVIDIA GeForce 940MX
        Vendor     : NVIDIA Corporation
        Driver     : CUDA 12.2
        Aspects    : gpu fp64 online_compiler online_linker queue_profiling usm_device_allocations usm_host_allocations usm_shared_allocations ext_intel_pci_address ext_intel_device_info_uuid ext_intel_free_memory ext_intel_device_id ext_intel_memory_clock_rate ext_intel_memory_bus_widthImages are not fully supported by the CUDA BE, their support is disabled by default. Their partial support can be activated by setting SYCL_PI_CUDA_ENABLE_IMAGE_SUPPORT environment variable at runtime.
 ext_oneapi_bindless_images ext_oneapi_bindless_images_shared_usm ext_oneapi_bindless_images_2d_usm ext_oneapi_interop_memory_import ext_oneapi_interop_semaphore_import ext_oneapi_mipmap ext_oneapi_mipmap_anisotropy ext_oneapi_mipmap_level_reference ext_oneapi_ballot_group ext_oneapi_fixed_size_group ext_oneapi_opportunistic_group ext_oneapi_graph ext_oneapi_limited_graph ext_oneapi_cubemap ext_oneapi_cubemap_seamless_filtering ext_oneapi_bindless_sampled_image_fetch_1d_usm ext_oneapi_bindless_sampled_image_fetch_2d_usm ext_oneapi_bindless_sampled_image_fetch_2d ext_oneapi_bindless_sampled_image_fetch_3d ext_oneapi_queue_profiling_tag
        info::device::sub_group_sizes: 32
default_selector()      : gpu, NVIDIA CUDA BACKEND, NVIDIA GeForce 940MX 5.0 [CUDA 12.2]
accelerator_selector()  : No device of requested type available. -1 (PI_ERRO...
cpu_selector()          : cpu, Intel(R) OpenCL, Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz OpenCL 3.0 (Build 0) [2025.20.6.0.04_224945]
gpu_selector()          : gpu, NVIDIA CUDA BACKEND, NVIDIA GeForce 940MX 5.0 [CUDA 12.2]
custom_selector(gpu)    : gpu, NVIDIA CUDA BACKEND, NVIDIA GeForce 940MX 5.0 [CUDA 12.2]
custom_selector(cpu)    : cpu, Intel(R) OpenCL, Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz OpenCL 3.0 (Build 0) [2025.20.6.0.04_224945]
custom_selector(acc)    : No device of requested type available. -1 (PI_ERRO...

Additional context

Trying to compile llama.cpp SYCL backend and have it recognize and use my GPU (ggml-org/llama.cpp#10929 (comment))

[AlexeySachkov]

Hi @drzraf, thanks for the report, even though I think that the behavior you see might be expected

The iGPU is wrongly seen as a CPU. The additional GeForce is found.

It is technically impossible for iGPU to be seen as CPU. SYCL is an abstraction layer on top of other runtimes such as OpenCL or CUDA. What sycl-ls shows is which low-level runtimes are installed and configured on your system which can communicate with the HW you have.

From the output it seems like you have CUDA installed and two OpenCL runtimes installed both of which can work with your CPU, but you don't have an OpenCL GPU runtime installed (or present in LD_LIBRARY_PATH).

You can obtain the latter from intel/compute-runtime. Please note that your GPU is considered to be as a legacy platform (docs), i.e. the latest driver may not support it properly anymore.

About the difference between official oneAPI packages and intel/llvm: we have SYCL_DEVICE_ALLOWLIST env variable which allows us to say to SYCL RT to ignore some low-level runtimes. oneAPI releases come with sycl.conf file in the distribution which automatically presets that env variable through the configuration file. See also #17114

We do that to exclude some platforms that we know are either unstable or untested. For example for CPU the filter is set to only show Intel's OpenCL implementations of certain versions, so POCL is ignored by default. intel/llvm does not have any prepopulated sycl.conf, so SYCL RT reports everything it found including POCL.

As for CUDA devices, oneAPI installation does not come with CUDA & HIP adapters out of the box, so SYCL RT is not capable of detecting those devices at all. You need to install those plugins separately from https://developer.codeplay.com/products/oneapi/nvidia/download

[drzraf]

Thank you for the great explanation (I was really lost between ONEAPI_DEVICE_SELECTOR and SYCL_DEVICE_ALLOWLIST and would have never guessed about this one.)

The key was:

sycl-ls shows is which low-level runtimes are installed and configured on your system

Installing intel-opencl-icd fixed that detection problem

[opencl:cpu][opencl:0] Intel(R) OpenCL, Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz OpenCL 3.0 (Build 0) [2025.20.6.0.04_224945]
[opencl:gpu][opencl:1] Intel(R) OpenCL Graphics, Intel(R) HD Graphics 620 OpenCL 3.0 NEO [23.43.027642]

With this I could further attempt to offload layers onto this GPU... but it's still failing

Thank you for the support though !

[KornevNikita]

@drzraf hi, so, was the issue addressed? Can we close this one or do you want to discuss something else here?