Compatibility with libsycl.so.6

It seems the conda package on channel dppy/label/dev is compatible with libsycl.so.5 but not libsycl.so.6. But intel just released 2023.0.0 versions of all the dependencies on conda and pypi and it ships libsycl.so.6 thus breaking the dpctl conda distribution:

    from ._sycl_device import SyclDevice, SyclDeviceCreationError
ImportError: libsycl.so.5: cannot open shared object file: No such file or directory

Would it be possible to update the build to work with libsycl.so.6 ?

Maybe dpnp and numba_dpex in the same channel will be affected too ?

In the meantime the dpcpp runtime can be pinned in conda, e.g with

export CONDA_DPEX_ENV_NAME=my-dpex-env
conda create --yes --name $CONDA_DPEX_ENV_NAME \
             --channel dppy/label/dev \
             --channel conda-forge \
             --channel intel \
             `# TODO: remove <2023.0.0 pin after dppy/label/dev is updated` \
             numba-dpex "intel::dpcpp_linux-64<2023.0.0"

dpctl 0.14.1dev1 built with DPC++ 2023.0.0 has been uploaded to dppy/label/dev. It would be compatible with libsycl.so.6

Thank you for the bumps.

I'm getting this new error though:

AttributeError: 'dpctl._sycl_device.SyclDevice' object has no attribute 'sub_group_sizes'

(using the custom build from this Dockerfile)

When on CPU (SYCL_DEVICE_FILTER=cpu) I'm also having dpctl._sycl_queue.SyclKernelSubmitError: Kernel submission to Sycl queue failed. errors. How can I debug such errors ? (reported in #1023)

Navigation Menu

Repository navigation

Compatibility with `libsycl.so.6` #1022

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Compatibility with libsycl.so.6 #1022

Description

Activity

fcharras commented on Dec 19, 2022

fcharras commented on Dec 19, 2022

oleksandr-pavlyk commented on Dec 27, 2022

fcharras commented on Dec 28, 2022

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Compatibility with `libsycl.so.6` #1022