[CPU] [ARM] SVE FP16 functions for MHASingleToken kernel

[NishantPrabhuFujitsu]

Details

In continuation with #27273, adds SVE FP16 implementations for functions called during execution of MHASingleToken for SVE-128, SVE-256 and SVE-512 platforms. SVE implementations are compiled only if runtime support for SVE is detected on the hardware, otherwise it falls back to Neon.

Benchmarking results

Below are the benchmarking results of execution time of each ported function. Measurements were performed by running each function individually on dummy inputs (128 fp16 elements) for 1,000,000 iterations and computing average time (in micro-seconds).

[NishantPrabhuFujitsu]

@dmitry-gorokhov you might remember a conversation we had about the inference output not being correct when the new exp_ps_sve implementation I added in #27273 is used in exp_reduce_sum. This issue is no longer happening, i.e. inference is working fine with our new exp(x) implementation.

I am not 100% sure what change caused this to get fixed but I will try to find out more soon.

[dmitry-gorokhov]

build_jenkins

[NishantPrabhuFujitsu]

I'm not very sure if the fix I just pushed for https://github.com/openvinotoolkit/openvino/actions/runs/12465345215/job/34792490178?pr=28182#step:15:3723 is correct. Could you please check if it would work?

[dmitry-gorokhov]

build_jenkins

[dmitry-gorokhov]

So we have Android ARM build failures. It happens because Android host uses compiler that actually doesn't support SVE, but we are still trying to cross-compile MHA for SVE isa.
To fix this we need to adjust ENABLE_SVE option based on ov_check_compiler_supports_sve check. I think we should add such check here https://github.com/openvinotoolkit/openvino/blob/master/cmake/developer_package/features.cmake#L108 (as it done for avx512).

[NishantPrabhuFujitsu]

I've added compiler version checks for clang (>13, source) and gcc (>10, source) in features.cmake. I couldn't find information about MSVC, and SVE is not supported on Apple platforms anyway. Hopefully this will be sufficient.

For the other failing pipelines, I remember those being CI issues encountered in #27273 as well. I hope that's still the case.

[dmitry-gorokhov]

I've added compiler version checks for clang (>13, source) and gcc (>10, source) in features.cmake. I couldn't find information about MSVC, and SVE is not supported on Apple platforms anyway. Hopefully this will be sufficient.

@NishantPrabhuFujitsu As I mentioned instead of compiler versions check it is more sufficient to implement condition based on ov_check_compiler_supports_sve call. ov_check_compiler_supports_sve tries to compile SVE code with given compiler - that seems to be most reliable way to check if SVE is supported.

[NishantPrabhuFujitsu]

@dmitry-gorokhov In that case, I'm not sure I understand what I need to change.

For Android platforms, ov_check_compiler_supports_sve check already seems to work as expected: CXX_HAS_SVE check fails (see CI logs, I've taken this from #27841 but I remember seeing the same for this PR) and the next two lines state that SVE is not supported on the platform. Isn't that enough to prevent SVE from being cross-compiled on Android?

If my understanding is wrong, please help understand how exactly this works and what adjustment to the check would fix it. Thanks!