[CPU] add arithmetic_mode impl for bf16_emitters

[liubo-intel]

Details:

• add arithmetic_mode impl for bf16_emitters to fix the 'inf' out issue when input data is out of rang [bf16_min,bf16_max] during f32->bf16, which may lead to 'UNK' outputs in some LLM cases

Tickets:

• CVS-157254

[liubo-intel]

Hi, @chenhu-wang and @dmitry-gorokhov : could you please help review this pr ? thanks

[chenhu-wang]

This is not a truncation behavior, it is saturation implementation by clap with [min, max], but named truncated.

[liubo-intel]

Hi, @chenhu-wang : thanks for your comments, do you suggest use some kind of more efficient instruction set to solve this kind of problem?

[chenhu-wang]

I mean "truncation" mode is to cut off lower bits directly. "saturation" mode is to clap with [min, max]. They have different result. PR description/variable name and this implementation is not aligned. First of all, we need understand which mode we needed here and align on all platforms. By truncation, you can ref h->uni_vandps(aux, aux, table_val("mask_truncation_word")); in below lines.

[liubo-intel]

I got your point, based on the customer model outputs, it seems we need this kind of "saturation" mode here.
and the new added functional tests also shows that "saturation" mode impl align with ov reference impl. Maybe I need to modify the PR description based on your suggestion.

[chenhu-wang]

@liubo-intel , You use fixed "saturation" mode on avx512_core_bf16 ISA, but on some ISA like avx2, fixed "truncation" mode is used. This is not aligned.
We have arithmetic_mode mode_ in jit_store_emitter. jit_uni_vcvtneps2bf16 is used in store f32 to bf16 case, where the mode info is not respect there. I think we should take mode_(truncation or saturation) into jit_uni_vcvtneps2bf16 as a field and convert according to the mode. We can set the mode "saturation" as default.