half float

[vinnitu]

Hello!

What about using half float in data_t?

http://half.sourceforge.net/

[yurymalkov]

Hi @vinnitu ,
Is it faster to compute compared to 32-bit float AVX?

[vinnitu]

maybe not faster but more compact footprint in memory I think

[VivianSnow]

@yurymalkov
https://software.intel.com/en-us/articles/performance-benefits-of-half-precision-floats
"Using half-floats provides a performance benefit over 32-bit floats when 32-bit float data does not fit into the L1 cache."

[vinnitu]

so, Yury we can do this? I can help you to test

[yurymalkov]

I can work on this, but somewhat later (too busy this non-programming world issues right now :( ).
If someone is willing to do this earlier, I will find some time to help.

[VivianSnow]

I can work on this, but what's the best practice for controlling it? Is that add a new member variable in HierarchicalNSW?

[yurymalkov]

I think the best way to do it is create a new space.
E.g. in a similar way to uint8 type: https://github.com/nmslib/hnswlib/blob/master/hnswlib/space_l2.h#L216