SYCL: do not skip CPU or FPGA if explicitly selected

[smeso]

It can be useful to be able to use SYCL on CPU
e.g. to run tests on a machine without a GPU.

With this commit a user can enable the use
of the CPU with the following env variable:

export ONEAPI_DEVICE_SELECTOR="opencl:cpu"

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

I ran the CI locally with this change and I got this error:

NORM(type=f32,ne=[64,10,10,10],eps=0.000001): /home/debian/llama.cpp/ggml/src/ggml-sycl/norm.cpp:11: void norm_f32(const
 float *, float *, const int, const float, const sycl::nd_item<3> &, sycl::float2 *, int): global id: [0,0,0], local id: [0,0,0] Assertion `nwarps % WARP_SIZE == 0` failed.

I guess this only happens because the code isn't supposed to run on a CPU at all and I'm running it on a CPU using this commit.

Would it be okay to make the code more generic so that it could also run on CPU? Or do you think that it would be a lot of work?

The main reason why I'm doing this is that I don't have access to a GPU that the current SYCL code can use and I had to make some changes to SYCL code because of ggerganov/ggml#940 and I didn't know how to test those changes.
Personally I don't have a problem if some tests fail on CPU, maybe this could be a "use at your own risk" mode?

[NeoZhangJianyu]

@smeso
Thank your comment!
We have considered SYCL support CPU.
But current SYCL code includes many hardcode for GPU performance.
To support CPU, lots of hardcode need be updated for CPU hardware.

Because there is always AVX backend to support CPU better, there is no value to make SYCL support CPU.

If you want to test SYCL locally, you could use iGPU in Intel 11th Core or newer CPU.
For Meteor Lake or newer CPU, the built-in Arc GPU is powerful.

[smeso]

Oh thanks for the reply!
I was indeed trying to use SYCL with and older generation CPU, I should try with a more recent one.
I'll close this PR.