Level Zero module creation fails

[diptorupd]

Encountering the following error with dpctl master anddpcpp 2021.4

Note: Google Test filter = TestQueueSubmit.CheckSubmitRange_saxpy
[==========] Running 1 test from 1 test suite.
[----------] Global test environment set-up.
[----------] 1 test from TestQueueSubmit
[ RUN      ] TestQueueSubmit.CheckSubmitRange_saxpy
    Name            Intel(R) UHD Graphics 630 [0x3e98]
    Driver version  1.2.20889
    Vendor          Intel(R) Corporation
    Profile         FULL_PROFILE
    Filter string   level_zero:gpu:0
The level zero loader dynamic library could not be opened.
ZeModuleCreateFn is invalid.

[diptorupd]

@oleksandr-pavlyk I seem to not have libze_loader.so on my system. I do have libze_loader.so.1 and libze_loader.so.1.4.1. Any idea what package I am missing?

@JacekDanecki any pointers as to how to resolve it? I have oneAPI 2021.4 installed. More details #609 (comment)

[oleksandr-pavlyk]

I encountered the same issue with OS LLVM/SYCL, see #603

I guess the library had been renamed.

[diptorupd]

or the libze-loader Ubuntu package is corrupted. I built the library by hand and do see libze_loader.so.

[diptorupd]

@oleksandr-pavlyk I have a temporary workaround: build and install libze_loader by hand (https://github.com/oneapi-src/level-zero) and then add the library to the LD_LIBRARY_PATH.

The compute runtime page suggests installing apt install libze-loader libze-intel-gpu which is what I have on my system. However, the oneapi dgpu installation guide suggests other packages:

sudo apt-get update
sudo apt-get install \
  intel-opencl-icd \
  intel-level-zero-gpu level-zero \
  intel-media-va-driver-non-free libmfx1

I cannot even install the intel-level-zero-gpu and level-zero packages without first having to install libigc1 and libigdfcl1.

@PokhodenkoSA @reazulhoque do you know how level zero loader is installed in our TeamCity systems? I wonder why we are not running into the issue in our internal CI. Maybe the drivers are not updated in a while?

[oleksandr-pavlyk]

I always uninstall all drivers, and then reinstall them using one command from install Intel GPU drivers page for apt on Ubuntu 20.04

I did that on the linux CI machine 2 weeks back.

[PokhodenkoSA]

@diptorupd
I faced this issue IntelPython/numba-dpex#530.
libze_loader.so is in apt package level-zero-dev.

[PokhodenkoSA]

I think dpctl should load dynamically libze_loader.so.1 instead of libze_loader.so.

[PokhodenkoSA]

For me I find this documentation not working: compute runtime page.
For rolling back I had to remove ppa: add-apt-repository --remove ppa:intel-opencl/intel-opencl.
After that I use instructions from: oneapi dgpu installation guide.

[diptorupd]

@diptorupd I faced this issue IntelPython/numba-dppy#530. libze_loader.so is in apt package level-zero-dev.

Yes, that worked for me.

[oleksandr-pavlyk]

I think dpctl should load dynamically libze_loader.so.1 instead of libze_loader.so.

We definitely should align with whatever ${CONDA_PREFIX}/lib/libpi_level_zero.so} links to:

(idp_2021.4) [15:31:52 ansatnuc04 dpctl-capi]$ ldd /opt/intel/oneapi/intelpython/python3.7/lib/libpi_level_zero.so | grep libze_
        libze_loader.so.1 => /lib/x86_64-linux-gnu/libze_loader.so.1 (0x00007f450870e000)

We should infer this filename at the build time, and inject this filename in the source before starting the build.

Perhaps make it a part of dpctl-capi/Config/dpctl_config.h.in.

[diptorupd]

closing as installing level-zero-dev solves the problem.