Efficient AVX512 implementation in 'InnerProductSIMD16ExtAVX512' Function

[aurora327]

InnerProductSIMD16ExtAVX512 functions are implemented using the more efficient AVX512 instruction set

[aurora327]

Hi @yurymalko, Can you please to review the code?

[yurymalkov]

Hi @aurora327,

Thank you for the PR! I am slow to respond currently due to sickness, sorry.

I wonder, how much improvement did you see in the tests with the better implementation?

[aurora327]

hi, @yurymalkov
Different Size of vector1 and vector2 of the passed parameters have different performance gains, my own dataset build on 4th Generation Intel® Xeon® resulted in a 2% to 10% end-to-end improvement with 4 cores bound.
I hope you feel better soon :)

[yurymalkov]

Thanks again for the PR! I've also checked the query performance, it is up to 15% for 16-dim.
I also wonder if aligned/unaligned memory makes a difference for the current architectures?

[aurora327]

No differences were observed since the avx512 instruction used handles both aligned and unaligned data well. At the same time, the aligned buffer is highly recommended if possible.