oneDNN: CUDA backend: Application couldn't find GPU

Version

v2.5

Environment:

cpu: Intel® Xeon® Gold 6230 CPU @ 2.10GHz gpu: Tesla V100-PCIE-32GB os: CentOS 7.9.2009 gcc: 8.3.0 cuda: 10.2 cmake: 3.16.3 clang: intel/llvm@8458fab opencl runtime: oneapi_21.4

Step to reproduce

Same as #1215

Test

$ cd build 
$ make test 
99% tests passed, 1 tests failed out of 118

Total Test time (real) = 396.53 sec

The following tests FAILED:
	 26 - test_deconvolution_buffer_gpu (Failed)
Errors while running CTest
make: *** [test] Error 8

Examples

$ clang++ -fsycl -fsycl-targets=nvptx64-nvidia-cuda getting_started.cpp
$ ./a.out cpu 
Example passed on CPU.
$ ./a.out gpu
Application couldn't find GPU, please run with CPU instead.

Issues

• I am getting conflicted result when running post-build tests and getting started example. The former passed all tests except the deconvolution one. Yet the later cannot initialize the GPU.
• Since I cannot modify icd file located in /etc/OpenCL/vendors/, I had to use the runtime from oneAPI 2021.4 release. This allows clinfo to ‘see’ both CPU and GPU platforms.
• I think explicit GPU selection is required for CUDA backend in getting_started.cpp. Could you give me some pointers ?