[SYCL] Add oneDNN primitive support

[luoyu-intel]

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

oneMKL and oneDNN almost share the same GPU kernels, but oneDNN has better compatibility.

llama-bench on A750:

DNNL F16
| model                          |       size |     params | backend    | ngl |          test |              t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------------: | ---------------: |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  99 |        pp1024 |   1607.21 ± 5.46 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  99 |          tg32 |     32.01 ± 0.09 |

MKL F16
| model                          |       size |     params | backend    | ngl |          test |              t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------------: | ---------------: |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  99 |        pp1024 |   1581.72 ± 4.02 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  99 |          tg32 |     32.08 ± 0.13 |

DNNL F32
| model                          |       size |     params | backend    | ngl | threads |    sm |          test |              t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | ----: | ------------: | ---------------: |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  33 |       8 |  none |        pp1024 |    710.00 ± 0.42 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  33 |       8 |  none |          tg32 |     35.77 ± 0.02 |

MKL F32
| model                          |       size |     params | backend    | ngl | threads |    sm |          test |              t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | ----: | ------------: | ---------------: |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  33 |       8 |  none |        pp1024 |    709.28 ± 0.92 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  33 |       8 |  none |          tg32 |     35.74 ± 0.04 |

on A770m, the master branch crashed with GGML_SYCL_F16=ON, this branch with oneDNN can run it.

DNNL F16

| model                          |       size |     params | backend    | ngl | threads |    sm |          test |              t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | ----: | ------------: | ---------------: |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  33 |       8 |  none |        pp1024 |   1598.31 ± 1.73 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  33 |       8 |  none |         tg128 |     40.20 ± 0.09 |

Ultra-155H, MKL throws this error:

Native API failed. Native API returns: -2 (PI_ERROR_DEVICE_NOT_AVAILABLE) -2 (PI_ERROR_DEVICE_NOT_AVAILABLE)
Exception caught at file:C:/Users/luoyu/Documents/repo/llama-fork/ggml/src/ggml-sycl.cpp, line:2556, func:operator()
SYCL error: CHECK_TRY_ERROR(dpct::gemm( *stream, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, &alpha_f16, src0_ptr, dpct::library_data_t::real_half, ne00, src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16, dst_f16.get(), dpct::library_data_t::real_half, ldc, dpct::library_data_t::real_half)): Meet error in this line code!

DNNL F16
| model                          |       size |     params | backend    | ngl | threads |          test |              t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | ------------: | ---------------: |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  33 |       8 |        pp1024 |   309.47 ± 39.50 |
| llama 7B Q4_0                  |   3.56 GiB |     6.74 B | SYCL       |  33 |       8 |          tg32 |     11.67 ± 1.07 |

[NeoZhangJianyu]

It's great to involve the oneDNN in SYCL backend.

1. If oneDNN is detected, oneMKL won't be used in whole code.
   Is it right?

2. Please update the SYCL.md with oneMKL and oneDNN dependence info.

[luoyu-intel]

It's great to involve the oneDNN in SYCL backend.

1. If oneDNN is detected, oneMKL won't be used in whole code.
   Is it right?
2. Please update the SYCL.md with oneMKL and oneDNN dependence info.

1. Yes.
2. MD has been updated.

[AGenchev]

Interesting. I wonder whether this thing, similar to OneDNN https://github.com/amd/ZenDNN could be integrated into SYCL backend.

[vpirogov]

Interesting. I wonder whether this thing, similar to OneDNN https://github.com/amd/ZenDNN could be integrated into SYCL backend.

This thing is CPU only and is not intended to be used as standalone software. oneDNN will cover AMD client and server CPUs nicely if necessary.