Including queryNorms for runL2SelectMin?

[gauenk]

My question refer to file faiss/faiss/gpu/impl/Distances.cu lines 230 - 250.

A few notes to agree on:

• The query norms are not required to select a query's k nearest neighbors, since it is a constant for each "centroid"
• The "distanceBufView" object covers a large portion of memory
• If we want to compute the exact distances for all elements (ignoreOutDistances = False), then we must run over distanceBufView twice.
• Running over the distanceBufView is "slow"

If we agree with the above statements, I am prepared to ask my question: Why does runL2SelectMin not optionally include the queryNormNiew in its operation? I can understand how we might save some computation by removing an extra addition term. However, if this was the case the if-statement for "ignoreOutDistances" can just switch between the two versions of runL2SelectMin function: one with the queryNorms and one without them.

Okay, so people are busy and it's possible this optimization is just not worth people's time. I am wondering if this design was more deliberate.

I am thinking there is a more significant (maybe CUDA-related) reason to separate the queryNorms addition? Specifically, say we include the additional vector "queryNorm" in the "l2SelectMinK" function in the gpu/impl/L2Select.cu file. Maybe the inclusion of an additional vector impacts hardware utilization of warps, since they operate under some locality assumptions? By including this additional vector (like queryNorms), we might overflow some mysterious, behind-the-scenes CUDA-cache? I am far from a CUDA expert, so I am merely guessing here.

Any insight would be helpful. Thank you in advance.