SYCL : Move to compile time oneMKL interface backend selection for NVIDIA backend

[s-Nick]

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

This patch move oneMKL interface calls to gemm from real time to compile time for NVIDIA backend, bringing improvements specially in text generation.

Tested on A100:
Current

model	size	params	backend	ngl	sm	mmap	test	t/s
llama 70B Q4_K - Small	37.57 GiB	70.55 B	SYCL	99	none	0	pp512	705.51 ± 2.23
llama 70B Q4_K - Small	37.57 GiB	70.55 B	SYCL	99	none	0	tg128	14.28 ± 0.05
llama 8B Q4_K - Medium	4.58 GiB	8.03 B	SYCL	99	none	0	pp512	5426.17 ± 29.64
llama 8B Q4_K - Medium	4.58 GiB	8.03 B	SYCL	99	none	0	tg128	81.36 ± 1.29
llama 8B Q8_0	7.95 GiB	8.03 B	SYCL	99	none	0	pp512	5592.87 ± 89.22
llama 8B Q8_0	7.95 GiB	8.03 B	SYCL	99	none	0	tg128	72.96 ± 0.91

build: 0f77aae (20)

With changes

model	size	params	backend	ngl	threads	sm	mmap	test	t/s
llama 70B Q4_K - Small	37.57 GiB	70.55 B	SYCL	99	8	none	0	pp512	720.68 ± 1.62
llama 70B Q4_K - Small	37.57 GiB	70.55 B	SYCL	99	8	none	0	tg128	18.52 ± 0.07
llama 8B Q4_K - Medium	4.58 GiB	8.03 B	SYCL	99	8	none	0	pp512	5489.17 ± 30.44
llama 8B Q4_K - Medium	4.58 GiB	8.03 B	SYCL	99	8	none	0	tg128	91.99 ± 0.05
llama 8B Q8_0	7.95 GiB	8.03 B	SYCL	99	8	none	0	pp512	5439.13 ± 216.28
llama 8B Q8_0	7.95 GiB	8.03 B	SYCL	99	8	none	0	tg128	89.34 ± 0.05

build: ffd0a99 (4222)

[s-Nick]

@Alcpz Could you check it out?

[Alcpz]

@Rbiessy Feel free to give it a look, since you have experience working with oneMKL interface

[NeoZhangJianyu]

Same comments for other update.

[NeoZhangJianyu]

The macro make the code is hard to understand.
I suggest:

#ifdef GGML_SYCL_NVIDIA
        oneapi::mkl::blas::column_major::gemm(
                oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q },
                a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
                data_b, ldb, beta_value, data_c, ldc);
        }
#else
        oneapi::mkl::blas::column_major::gemm(
                q, 
                a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
                data_b, ldb, beta_value, data_c, ldc);
        }
#endif

[Rbiessy]

If we start adding support for Intel GPU as well I think it would make more sense to have a helper function that returns either a backend_selector or a queue based on the backend.
It would avoid duplicating the call to gemm which I think is a risk.

[NeoZhangJianyu]

Please remember, the SYCL backend is initiated to support Intel GPU. :)
Support more vendor GPUs is added later.
The default code path should be optimized for Intel GPU.

It's OK to set special queue for other vendor GPUs.

[s-Nick]

Code update for readability in f6e6fc4

[NeoZhangJianyu]

@s-Nick
I guess the same method could help for Intel GPU.
Is it possible to test for Intel GPU too? like oneapi::mkl::backend::mklgpu.

[s-Nick]

Thank you for your review @NeoZhangJianyu
Currently Intel GPU implementation uses oneMKL closed source library directly and it doesn't have nor need a backend_selector, therefore these changes aren't required or useful

[Rbiessy]

The oneMKL Interface changes look good to me.

[NeoZhangJianyu]

Thank you for your review @NeoZhangJianyu Currently Intel GPU implementation uses oneMKL closed source library directly and it doesn't have nor need a backend_selector, therefore these changes aren't required or useful

OK, I see!

[Alcpz]

changes lgtm. Let's wait for the remaining thread to be resolved before merging.