run_prebuilt_e2e_tests CI jobs fail in cases with UR API changes

[EwanC]

As discovered by @yingcong-wu in #16747 (comment) the run_prebuilt_e2e_tests CI jobs in PR testing fail in the case that the UR API changes.

I notice that those features are all tested in llvm/sycl/test-e2e/lit.cfg.py using subprocess.getstatusoutput. I added a print to those tests and they all show the following error:

2025-02-11T15:09:01.2540004Z lit.py: /__w/llvm/llvm/llvm/sycl/test-e2e/lit.cfg.py:328: note: /usr/bin/ld: /__w/llvm/llvm/toolchain/bin/../lib/libsycl.so: undefined reference to `urEnqueueCommandBufferExp@LIBUR_LOADER_0.12'
2025-02-11T15:09:01.2540580Z clang++: error: linker command failed with exit code 1 (use -v to see invocation)
2025-02-11T15:09:01.2541173Z lit.py: /__w/llvm/llvm/llvm/sycl/test-e2e/lit.cfg.py:356: note: /usr/bin/ld: /__w/llvm/llvm/toolchain/bin/../lib/libsycl-preview.so: undefined reference to `urEnqueueCommandBufferExp@LIBUR_LOADER_0.12'

(In https://github.com/intel/llvm/actions/runs/13265484428/job/37031762080, line 104607).

I think there is a problem with those feature checks(maybe linking to an out-dated UR libs?), and I think it would impact any UR ABI breaking changes.

Experimental features in Unified Runtime are allowed to make API/ABI breaking changes, and the SYCL-Graph experimental oneAPI extension relies on being able to do this. This CI job should be updated so that it can pass in such a case, or a policy should be established with the gatekeepers whereby PRs that fail run_prebuilt_e2e_tests CI runs due to this reason can still be merged.

[sarnex]

@ayylol Do you mind taking a look at this? It seems specific to prebuilt e2e testing. If that's not related to the problem let me know and I'll investigate.

[aelovikov-intel]

AFAIK, normally one can use /path/to/freshly/built/clang++ -fsycl without setting extra env. We rely on that here:

llvm/.github/workflows/sycl-linux-build.yml

Line 281 in 0e44bb4

cxx_compiler: $GITHUB_WORKSPACE/toolchain/bin/clang++

. Did UR break that? @EwanC

[ayylol]

Hey @EwanC, for some context the way the "dev_kit" features are added, which seems to be the ones causing the issues in your pr, is by testing to see if a simple file that depends on those dev_kits can be compiled. In the case for the level_zero_dev_kit we try to compile the following file:

#include <level_zero/ze_api.h>
int main() { uint32_t t; zeDriverGet(&t, nullptr); return t; }

using the following command:

clang++ -fsycl l0_include.cpp -L<path-to-repo>/build/lib -lze_loader  -I<path-to-repo>/build/_deps/level-zero-loader-src/include

Given the similarity of the error message undefined reference to 'urEnqueueCommandBufferExp@LIBUR_LOADER_0.12' and the pr title "Update to UR cmd-buf enqueue rename" I don't think its a CI issue, and rather a UR issue from that commit.

[EwanC]

Thanks for investigating, I've tried that reproducer locally using #16984 which is the same change and shows the same CI issues in the prebuilt e2e test. Using the reproducer you suggest seems to work fine locally. Is there something I need to tweak locally to reproduce this issue?

$ git status
On branch ewan/rename_enqueue
$ cd build
$ ninja deploy-sycl-toolchain
$ cat l0_include.cpp
#include <level_zero/ze_api.h>
int main() { uint32_t t; zeDriverGet(&t, nullptr); return t; }
$ ./bin/clang++ -fsycl l0_include.cpp -L $PWD/lib -lze_loader  -I $PWD/_deps/level-zero-loader-src/include -o /tmp/l0_include

$ LD_LIBRARY_PATH=$PWD/lib /tmp/l0_include
$ echo $?
72

This is the same output I see as tip sycl

$ git checkout sycl
Switched to branch 'sycl'
Your branch is up-to-date with 'github/sycl'.
$ cd build
$ ninja deploy-sycl-toolchain
$ ./bin/clang++ -fsycl l0_include.cpp -L $PWD/lib -lze_loader  -I $PWD/_deps/level-zero-loader-src/include -o /tmp/l0_include
$ LD_LIBRARY_PATH=$PWD/lib /tmp/l0_include
$ echo $?
72