Recommended building and runtime dependences for MKL.

[fengyuan14]

🚀 The feature, motivation and pitch

So far, #526 implemented the first MKL related operator, aten.fft_c2c. The PR introduced MKL building system in torch-xpu-ops. The MKL SDK introduced for building bases on oneAPI package.

The potential issue is we would recommend using pip install mkl-dpcpp for runtime. There would be potential API breaking issue when MKL version in oneAPI package for building has gap with MKL in Pypi.

We need to unify recommended MKL package for building and runtime.

Alternatives

No response

Additional context

No response

[gajanan-choudhary]

Is there a plan to add other oneMKL APIs such as those from Sparse BLAS in torch-xpu-ops? I'm seeing things like #1330 which appear to start adding Sparse CSR support but aren't (or maybe can't yet be?) using oneMKL.

[CuiYifeng]

@gajanan-choudhary Thanks for your question!
Some sparse ops, such as _sparse_sparse_matmul and _sparse_addmm, will be implemented with oneMKL.
Since ops in #1330 are implemented without MKL in Pytorch CPU, this choice has been retained. For now, functionality first, so relying less is better. These ops may be refined using oneMKL in future.