Closed
Description
Hi,
I am trying to compile some MKL code and use the CUDA plugin interface (ee21d66). Only the sycl/cuda toolchain is installed on the machine so I'm pulling MKL from the oneAPI (beta05) we have installed. Seems either they're incompatible, or I'm doing something wrong.
...
/opt/intel/inteloneapi/mkl/2021.1-beta05/lib/intel64/libmkl_sycl.so: undefined reference to `cl::sycl::queue::submit_impl(std::function<void (cl::sycl::handler&)>)'
/opt/intel/inteloneapi/mkl/2021.1-beta05/lib/intel64/libmkl_sycl.so: undefined reference to `cl::sycl::detail::Scheduler::addHostAccessor(cl::sycl::detail::AccessorImp
lHost*)'
/opt/intel/inteloneapi/mkl/2021.1-beta05/lib/intel64/libmkl_sycl.so: undefined reference to `cl::sycl::detail::half_impl::half::half(float const&)'
/opt/intel/inteloneapi/mkl/2021.1-beta05/lib/intel64/libmkl_sycl.so: undefined reference to `cl::sycl::detail::pi::CallPi<_pi_result (*)(_pi_mem*), 184ul>::CallPi()'
/opt/intel/inteloneapi/mkl/2021.1-beta05/lib/intel64/libmkl_sycl.so: undefined reference to `cl::sycl::detail::half_impl::half::operator float() const'
clang-11: error: linker command failed with exit code 1 (use -v to see invocation)
...
Should this work, or does MKL need a separate build for the CUDA PI? I note that the same code compiles fine if I do a straight dpcpp/sycl build.