Performance issue in IndexFlatIP

[Museum7432]

Summary

IndexFlatIP doesn't seem to use parallelization correctly, as my custom flat index is somehow 2x to 3x (depends on the number of threads) faster than IndexFlatIP when I wrote a benchmark to compare the two. I kind of expected mine to be slower than Faiss as it is the same under the hood with fewer optimizations, not faster like this.

Platform

OS: Linux

Faiss version: 1.9.0

Installed from: faiss-cpu from anaconda and pip

Running on:

• CPU
• GPU

Interface:

• C++
• Python

Reproduction instructions

Here is the colab notebook for reproducing the problem.
It could be caused by the memory access pattern in IndexFlatIP (like why does the function fvec_inner_products_by_idx parallelize over the queries not ids, I can't read the code for dispatch_knn_ResultHandler though, so i can't say it is caused by it)