[CPU] Fix avx2 bf16 reorder

[maxnick]

Details:

Fix forked oneDNN JIT reorder kernel for the case of bf16 precision and AVX2 ISA.

OneDNN fork PR: openvinotoolkit/oneDNN#253

Tickets:

• CVS-145874

[maxnick]

@luweizhou2016 , could you please review?

[luweizhou2016]

@maxnick , My understanding is that for avx2 plaform:
bf16->fp32 is still supported in jit_uni_reorder_kernel_f32_t;
fp32->bf16 is not supported in jit_uni_reorder_kernel_f32_t for the bf16_emu bug?

[maxnick]

@maxnick , My understanding is that for avx2 plaform: bf16->fp32 is still supported in jit_uni_reorder_kernel_f32_t; fp32->bf16 is not supported in jit_uni_reorder_kernel_f32_t for the bf16_emu bug?

fp32->bf16 is not supported by avx2 as the bf16_emu works only with avx512, and the avx2 instruction set doesn't have dedicated fp32->bf16 conversion instructions. Also fp32->bf16 reorders are supported by the avx2 platforms with the convert NE extension, and of course for AVX512_BF16, as they have native fp32->bf16 conversion instructions.

[luweizhou2016]

@maxnick , My understanding is that for avx2 plaform: bf16->fp32 is still supported in jit_uni_reorder_kernel_f32_t; fp32->bf16 is not supported in jit_uni_reorder_kernel_f32_t for the bf16_emu bug?

fp32->bf16 is not supported by avx2 as the bf16_emu works only with avx512, and the avx2 instruction set doesn't have dedicated fp32->bf16 conversion instructions. Also fp32->bf16 reorders are supported by the avx2 platforms with the convert NE extension, and of course for AVX512_BF16, as they have native fp32->bf16 conversion instructions.

Thx for clarifying. Approved.